Internet-Draft Context Management Architecture October 2026
Moussa & Akhavain Expires 1 May 2027 [Page]
Workgroup:
agentproto
Internet-Draft:
draft-agentproto-context-management-architecture-00
Published:
Intended Status:
Informational
Expires:
Authors:
H. Moussa
Huawei Canada
A. Akhavain
Huawei Canada

Context and Correlation Management Architecture for Distributed Agentic Activities

Abstract

Agentic applications increasingly distribute a single objective across autonomous agents, tools, and services implemented by different vendors, operated by different organizations, and connected through different protocols.

A single activity may contain multiple pairwise dialogs, multiple tasks within each dialog, tool invocations, derived resources, downstream delegations, and later interactions that reinterpret or reuse artifacts produced earlier in the activity. Existing protocols provide useful local identifiers for their own interactions, but those identifiers do not by themselves provide a common information model for correlating independently maintained dialogs, tasks, and resources across protocol and administrative boundaries.

This document defines an information model consisting of Activities, Dialogs, Tasks, Resources, and typed Relationships. It distinguishes Activity correlation from protocol-local Dialog context, and distinguishes stable resource identity and provenance from locally defined interpretations.

As such, this document develops a context-and-correlation information-model architecture, not a functional reference architecture or a wire protocol. It identifies relevant objects and relationships, relevant to maintaining coherent context, and examines the implications for preserving and propagating their correlations. It does not define the procedures by which applications execute work. The model distinguishes the logical extent of an Activity, which may extend beyond protocol-managed correlation; extending that coverage is selective, while maintaining established managed correlations entails responsibilities.

Status of This Memo

This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.

Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.

Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."

This Internet-Draft will expire on 1 May 2027.

▲

Table of Contents

1. Introduction

Distributed agentic systems are increasingly composed of multiple independently operated participants.

A single Activity can involve:

No single participant, protocol, or identifier spans all of these interactions.

Multiple such Activities may proceed concurrently while sharing the same underlying distributed agentic system, giving rise to multiple intersecting interaction waves composed of Dialogs, Tasks, conversations, artifacts, Resources, and other related objects. These waves may overlap, share participants, pursue competing objectives, and serve different users and goals. Under these circumstances, the challenge is to ensure that such waves can propagate across the system without causing significant confusion, loss, dropped interactions, or delays in service. In essence, the objective is to preserve sufficient correlation for independently maintained Dialogs, Tasks, and Resources to remain meaningfully related, such that value of distributed agentic networks can be materialized.

Existing protocols provide useful local identifiers for their own interactions, but those identifiers cannot, by themselves, correlate independently maintained objects across Dialogs and administrative domains. This document develops an information model for that problem and examines the correlations that need to be understood, preserved, or propagated across those boundaries.

This information-model architecture has a different purpose from a functional reference architecture that positions discovery, identity, authentication, authorization, dialog management, and transport as building blocks of agentic communications. This document concentrates on what must be correlated among Activities, Dialogs, Tasks, and Resources, which are objects defined in this document as part of the architecture, and on how those correlations can remain coherent across independently maintained interactions. These concepts support reasoning about dialog context; their inclusion does not, by itself, require each concept or relationship to be standardized as a protocol object.

An Activity can extend beyond the correlations maintained by any one protocol. Participants may choose whether to establish correlation with additional Dialogs, but those that undertake to maintain a correlation within a protocol-managed scope are responsible for preserving its coherence as relevant information changes. A failure to maintain such an established correlation differs from an undisclosed interaction that was never included in that protocol's managed scope.

The remainder of this document is organized as follows. Section 2 states the core architectural principle, and Section 3 defines the terminology used throughout the document. Section 4 introduces the architectural model and illustrates it through motivating examples that cross organizational boundaries. Section 5 examines how multiple independently maintained Dialogs can coexist within a single Activity, and Section 6 zooms in to the Tasks within a Dialog. Section 7 introduces correlation context and the implications of this architecture for its propagation. Section 8 describes the failure modes that arise from incorrect or insufficient propagation. Sections 9 to 16 present a worked propagation example, the relationship taxonomy, privacy considerations across administrative boundaries, and the relationship of this architecture to existing protocols and related work. Sections 17 to 19 conclude with architectural requirements, out-of-scope items, and open questions.

This document distinguishes settled architectural principles from open design questions. The open questions are deliberate: they identify the decision space reserved for the dialog-management protocol and for coordination with related work. They are candidates for resolution in subsequent protocol documents, not gaps in the architecture.

2. Core Architectural Principle

The architecture rests on a single central principle:

The model uses five primary objects:

Activity
Dialog
Task
Resource
Relationship

The model also defines four participant roles:

Activity Initiating Entity (AIE)
First Next Entity (FNE)
Assistant Agent (AA)
Assistant Tool (AT)

Roles are relative to the invocation chain: the AIE initiates an Activity, its directly invoked entity is the FNE, and entities invoked downstream are Assistant Agents or Assistant Tools. Terminology follows below.

This information model complements protocol-local context management mechanisms. A dialog-management protocol may use the model's semantics and convey relevant Correlation Metadata through bindings to existing protocols; this architecture does not define those bindings.

3. Terminology

4. Architectural Model

4.1. Architecture Visualization

An Activity begins when an Activity Initiating Entity (AIE) invokes one First Next Entity (FNE) to pursue its objective. The invocation establishes the Activity and its initial Dialog between the AIE and the FNE:

          Activity
    +-----------------------------+
    |                             |
    | +--+--+   Dialog   +--+--+  |
    | | AIE |<---------->| FNE |  |
    | +-----+            +-----+  |
    |                             |
    +-----------------------------+

If the goal of the Activity can be fulfilled by the FNE itself, no additional entities are invoked for that Activity. Alternatively, if assistance from additional entities is required, the FNE, or any downstream assistant entity, may invoke assistant entities. Dialogs are created between the invoking entity and each invoked assistant entity, and the respective Tasks are handled within their associated dialogs, all under the umbrella of the Activity, and the interaction wave propagates. Dialogs created in this way belong to the Activity but might not directly involve the AIE; each is related to the Dialog that created it through lineage relationships (see "Dialog Lineage"). The interaction wave may then look like the following:

          Activity
+------------------------------------------------------------------+
|                                                                  |
| +-----+   Dialog   +-----+   Dialog   +-----+   Dialog   +-----+ |
| | AIE |<---------->| FNE |<---------->| AA1 |<---------->| AA3 | |
| +-----+   task(s)  +-----+   task(s)  +-----+   task(s)  +-----+ |
|                       +                  +                       |
|                Dialog |           Dialog |                       |
|                       | task(s)          | task(s)               |
|                       v                  v                       |
|                    +-----+            +-----+                    |
|                    | AT1 |            | AT2 |                    |
|                    +-----+            +-----+                    |
|                                                                  |
+------------------------------------------------------------------+

An Activity groups multiple independently maintained Dialogs, each of which may contain Tasks:

Conceptually, the model looks like the following:

                   Activity
                      |
          +-----------+-----------+
          |                       |
        Dialog                  Dialog
          |                       |
        Tasks                   Tasks
       (EVENTS)                (EVENTS)
          |                       |
        +---------------------------+
        |     Resource Layer        |
        +---------------------------+

The figure above introduces the Resource Layer, the logical location where Resources pertaining to an Activity reside. The Resource Layer makes those Resources accessible to the Dialogs (and their tasks) of the Activity, and the same Resource may be used by more than one Dialog. For example, multimedia components may be consumed, produced, or transformed by multiple Dialogs within the same Activity. The Resource Layer is a logical construct: it does not prescribe where Resources are physically stored and does not imply a globally visible or centrally managed store.

Cross-Dialog lineage and shared Resource use are the architectural focus. The interaction graph need not be globally stored or visible; participants may retain only the correlations their operations require.

4.2. Motivating Multimedia Example

Under the above model, various use cases can be envisioned. As a simple motivating example, consider a user who wants a video V1 analyzed. The user, acting as the Activity Initiating Entity (AIE), starts an Activity with the objective "analyze video V1" and invokes Agent A. Agent A is the First Next Entity (FNE) of the Activity: the Dialog between the user and Agent A (D1) is established which extends the activity; a task under this dialog is created and used to assign the objective of the activity to Agent A over D1, and V1 becomes available to the Activity and D1 as an initial Resource.

Agent A decomposes the analysis task: The visual content requires a video-analysis specialist, and the audio content must be transcribed into text before it can be analyzed. Agent A therefore invokes Agent B, an Assistant Agent (AA) that can analyze video and text but not audio. A Dialog (D2) is created between Agent A and Agent B; Agent A transfers V1 to Agent B within D2 and assigns the video-analysis Task. Agent A also invokes Tool X, an Assistant Tool (AT), to transcribe the audio track of V1. A Dialog (D3) is created between Agent A and Tool X. Interaction with Tool X is one-shot in nature: Agent A submits the transcription Task and receives transcript T1 in response. The Task in D3 references V1 by its Resource identity; Tool X accesses V1 through the Resource Layer of the Activity. Agent A then provides T1 to Agent B within D2, where it is analyzed as text alongside the video.

Mapped to the participant roles:

User     AIE
Agent A  FNE
Agent B  AA  (invoked by Agent A)
Tool X   AT  (invoked by Agent A)

Note that roles follow the invocation chain rather than the nature of the entity: had the user invoked Tool X directly, for example, to pre-transcribe V1 before engaging Agent A, Tool X would have been an FNE rather than an AT.

The scenario is depicted below:

                     User (AIE)
                          |
                          | Dialog D1:
                          | (Activity, V1)
                          v
                   Agent A (FNE)
              +-----------+-----------+
              |                       |
   Dialog D2: |                       | Dialog D3:
  (V1, Tasks) |                       | (Task, V1 ref.)
              v                       v
        Agent B (AA)            Tool X (AT)
       video analysis       audio transcription
              ^                       |
              |                       v
              |                      T1
              |                       | (back to agent A)
              |                       |
              |                       |
              +------ Dialog D2 ----Agent A (FNE)
              (T1 provided to Agent B)

V1 and T1 pertain to the same Activity and reside in its Resource Layer. The example exhibits both cross-Dialog resource patterns of the Architectural Model: the same Resource, V1, is used by two Dialogs (D2 and D3), and T1 is produced in one Dialog (D3) and consumed in another (D2). It also illustrates the two ways a Resource can reach a participant. A Resource may be transferred to the receiving participant within a Dialog, as V1 is transferred to Agent B in D2, or it may be referenced by its stable Resource identity and accessed through the Resource Layer, as V1 is accessed by Tool X. The architecture does not mandate either mechanism. In both cases, the Resource retains a single logical identity, and, where needed, its provenance, so that every participant refers to the same Resource regardless of how it physically obtained it.

For Agent B to interpret T1 correctly, it needs to know that T1 is the transcript of the audio track of the same video V1 it is analyzing, and that both Resources belong to the same Activity. Essentially, we need the following relationship to be established and understood by Agent B:

T1 derived-from V1

A conversational identifier alone is insufficient to express these semantics. A protocol-local context identifier, such as an A2A contextId, groups the messages and Tasks of a single Dialog, but it cannot express that T1 was derived from V1, that T1 was produced in D3 and is meaningful to the Tasks of D2, or that D1, D2, and D3 all belong to the same Activity. Tool X's local job identifier is of no help either: it is meaningful only within D3. Moreover, Agent B may be concurrently analyzing similar videos for other Activities; without Activity correlation and Resource provenance, Agent B cannot distinguish this T1 from a transcript belonging to an unrelated Activity. What is required is a correlation model that carries Resource provenance and Activity correlation across Dialog boundaries. This architecture captures this correlation model as standardized typed Relationships, available across dialog boundaries.

4.3. Cross-Organization Operation

The example of the previous section becomes architecturally more interesting when its participants are operated by different organizations:

User (AIE)      invokes Agent A (Organization 1)
Agent A (FNE)   Organization 1
Agent B (AA)    Organization 2
Tool X (AT)     Organization 3

Dialog D2, between Agent A and Agent B, and Dialog D3, between Agent A and Tool X, cross organizational boundaries, while Dialog D1 involves only the user and Organization 1.

Each organization maintains its own local identifiers for the portions of the Activity it can see as shown below:

Organization 1:
  activity_id = O1-17

Organization 2:
  A2A context = AB-92
  A2A task = TB-44

Organization 3:
  tool job = JOB-991

In architectural terms, O1-17 is Organization 1's local handle for the Activity, AB-92 is the Protocol-Local Context of Dialog D2, TB-44 is a protocol-local Task identifier within D2, and JOB-991 is Tool X's local identifier for the transcription Task in D3.

These local identifiers need not be globally shared, unique, or even syntactically alike. Correlation must therefore work without replacing or exposing protocol-local state. The architecture separates:

logical correlation

from:

local identifier representation

A domain may maintain private mappings between the two.

The Resource Layer is logical, not administrative: V1 may be stored by Organization 1 and accessed by Tool X subject to that organization's access and privacy policies.

5. Dialogs Within an Activity

As described in the Architectural Model, a single Activity can involve multiple Dialogs.

An Activity does not imply a common conversational context, participant set, protocol instance, or lifecycle for its Dialogs. In other words:

Dialogs in different branches may have explicit relationships without direct communication or shared conversation. Common Activity membership alone does not imply a relationship.

An AIE may initiate multiple Activities to pursue a broader objective. The AIE may preserve associations between those Activities locally and coordinate their results or Resources across Activity boundaries. Consequently, Dialogs belonging to different Activities may be correlated indirectly through the AIE, without requiring a direct Dialog-to-Dialog relationship or disclosure of the other Activity to its participants. Having the same AIE does not, by itself, imply that those Activities are related.

This distinction supports delegation, specialization, cross-domain work, and recovery without collapsing independent interactions into one Dialog.

The motivating example illustrates this property: Dialogs D1, D2, and D3 are established by separate invocations, are maintained independently, and may use different application protocols, yet all belong to the same Activity.

5.1. Reasons for Creating a New Dialog

New Dialogs routinely arise as an interaction wave expands through delegation, specialization, tools, or recovery. For example:

  • Delegation: An agent may delegate part of its work to another agent.

  • Specialization: A participant may invoke a specialist because the current participants lack a required capability. For example, if Agent B cannot understand language L, it invokes an agent capable of language L and creates a new Dialog with it.

  • Tool invocation: Invoking a tool creates a separate Dialog for the one-shot interaction.

  • Administrative Separation: A new Dialog may be created when execution crosses into another organization, vendor, security domain, or policy boundary.

  • Failure Recovery: A replacement Dialog may be created after the failure of a participant, route to a participant, the transport, or an intermediary entity. For example, if Agent B fails mid-Activity, a replacement Agent B' is started and opens a new Dialog with Agent A to continue from where they stopped.

  • Privacy Separation: A participant may intentionally create a separate Dialog so that only a subset of Activity information is visible to the downstream participant. The same two participants may even run multiple Dialogs in parallel under the same Activity, each exposing a different subset of Activity information.

The motivating example already exhibits several of these reasons: D2 arises from delegating video analysis to a specialist, D3 from tool invocation, and both Dialogs cross administrative boundaries.

Applications decide when to establish Dialogs. The architectural concern is correlating each Dialog with its Activity and relevant lineage, where needed to capture causality.

5.2. Dialog Independence

Each Dialog is independently maintained.

In the motivating example, D2 and D3 both belong to the same Activity:

Activity: analyze video V1

D2:
  Agent A <-> Agent B

D3:
  Agent A <-> Tool X

D2 and D3 may differ in:

  • Participants: Agent B versus Tool X;

  • Protocol (e.g., HTTP vs MoQ) and transport (e.g., TCP vs UDP);

  • Protocol-local identifiers: A2A context AB-92 and Task TB-44 in D2, tool job JOB-991 in D3;

  • Resources referenced: V1 and T1 in D2, but V1 by reference in D3;

  • Message history;

  • Lifecycle state: D2 is long-running, D3 is one-shot;

  • Privacy policy and authorization level;

  • Administrative domain: Organization 2 versus Organization 3.

Therefore:

D2 != D3

even though both belong to the same Activity.

Dialog independence extends to the same pair of participants. Two entities may maintain multiple concurrent Dialogs with each other: one per Activity on which they both work, or several within a single Activity when privacy or task separation calls for it. Suppose that, alongside D2, Agent A opens a clarification dialog with Agent B:

Activity: analyze video V1

D2:
  Agent A <-> Agent B      (video-analysis Tasks)

D4:
  Agent A <-> Agent B      (clarification dialog)

Although D2 and D4 share participants, they have distinct local states and potentially different disclosures:

D2 != D4

A Dialog is identified by its own identity and the Activity it belongs to, never by its participant set.

A participant MUST NOT assume that another Dialog participating in the same Activity shares the same local state, and MUST NOT assume that its Dialog with a given counterpart is the only one.

5.3. Multi-Party Dialogs

A Dialog may involve more than two participants sharing one conversational context and lifecycle. For example, Agent A, Agent B, and Agent C may participate together:

    Single three-party dialog
  +-------------------------------+
  |            Agent A            |
  |               |               |
  |           +---+----+          |
  |          /          \         |
  |     Agent B        Agent C    |
  +-------------------------------+

This differs from Agent B invoking Agent C to create another Dialog:

Agent A <------> Agent B <------> Agent C
           D2               D5

D2 and D5 remain distinct when Agent A and Agent C do not share the same conversational context, even though D5 was created as a result of D2 with a relationship of for example:

D5 delegated-from D2.

The distinction rests on shared Dialog context, not on the number of interaction links between participants.

A multi-party Dialog's participant set may change over time:

Initially:       {Agent A, Agent B}
Agent C joins:   {Agent A, Agent B, Agent C}
Agent B leaves:  {Agent A, Agent C}

This architecture does not define participant-membership management semantics. A dialog-management protocol, however, may need to consider participant addition, removal, revocation, and membership changes as part of Dialog lifecycle management.

Finally, any multi-party Dialog can be decomposed into multiple pairwise Dialogs with appropriate relationship mappings. Whether to use one multi-party Dialog or several pairwise Dialogs is a design choice driven by privacy, cross-organization operation, and efficiency. A protocol for dialog management and context propagation SHOULD enable both modes of operation.

5.4. Dialog Lineage

Dialogs can be related through explicit lineage. For example:

D5 delegated-from D2

indicates that D5 exists because work from D2 was delegated into a new Dialog. A new Dialog may carry forward work or correlation context from a previous Dialog:

D2' continuation-of D2

Here D2 and D2' have distinct logical identities. This relationship does not represent resuming D2 with its original identity. A new Dialog may instead supersede an earlier Dialog:

D2' replaces D2

The two relationships describe different associations and may coexist. Their semantics are further discussed under "Recovery and Continuation".

This architecture distinguishes:

dialog identity

from:

dialog relationship

Two Dialogs remain independently identified even when one was created from another.

5.5. Hidden Delegation and Partial Visibility

A downstream Dialog need not be disclosed upstream. Suppose Agent B invokes Agent C:

Agent A <------> Agent B <------> Agent C
           D2               D5
             <------------->
                Delegation

Focusing on this delegation, the resulting visibility may be:

Agent A knows:
  Agent B
  D2

Agent B knows:
  Agent A
  Agent C
  D2
  D5
  D5 delegated-from D2

Agent C knows:
  Agent B
  D5

In the above example, Agent A does not know that Agent C exists. Agent C does not know that Agent A exists. In other words, the visibility level is a decision that participants make.

Such hidden delegation is common across organizational, vendor, trust, and privacy boundaries. The architecture separates:

relationship exists

from:

relationship is disclosed

A relationship can be retained locally by an intermediary without being propagated in full. For example, agents inside an organization need to know only that edge agent Y can help achieve objective X. Agent Y connects to agent Z in another organization; agent Z needs to know only that agent Y exists and is asking for V, not who asked Y. Yet agent Y must know that responses from agent Z relate to requests from internal agents.

5.6. Dialog Topologies

An Activity's Dialogs may branch, chain, and converge.

A participant may independently invoke another entity, so Activity topology MAY emerge from decentralized decisions.

The following scenarios are few possible examples of how topologies of an ACTIVITY MAY extend.

5.6.1. Branching

A participant in a Dialog can create several downstream Dialogs with other new participants as need arises. Agent B may invoke both Agent C and Agent E to pursue independent parts of its work:


         Activity A1
              |
           D1 |
              v     D2, D4
           Agent A <------> Agent B
              |               /\
           D3 |           D5 /  \ D6
              v             /    \
            Tool Y         v       v
                        Agent C   Agent E

with the relationships:

D5 delegated-from D2
D6 delegated-from D2

D5 and D6 are independent: they may execute concurrently, use different protocols, process different Resources, terminate independently, and expose different subsets of upstream context. Agent C and Agent E do not need to know of each other's existence.

5.6.2. Chaining

A downstream participant can itself create further Dialogs. If Agent C invokes Agent E, the Activity gains another level:


         Activity A1
              |
           D1 |
              v     D2, D4
           Agent A <------> Agent B
              |               /\
           D3 |           D5 /  \ D6
              v             /    \
            Tool Y         v      v
                        Agent C  Agent E
                           ^        ^
                           |   D7   |
                           +--------+

yielding:

D5 delegated-from D2
D7 delegated-from D5

The Activity therefore forms a delegation tree or, more generally, a graph. No participant is required to maintain the complete ancestry: Agent C may know that D5 has an opaque upstream origin without knowing the identity of Agent A or of D2.

5.6.3. Converging Dialogs

Separate branches can later contribute to another Dialog. If Agent B invokes an aggregator, Agent F, to combine the results produced independently in D5 and D6:

                D2
               /  \
              v    v
             D5    D6
              \    /
               \  /
                v
                D8

then D8 is correlated with more than one upstream Dialog:

D8 aggregated-from D5
D8 aggregated-from D6

aggregated-from indicates that D8 was created to aggregate or consume results produced independently in multiple upstream Dialogs. It is a multi-parent Dialog relationship, distinct from the single-parent delegated-from. Where only correlation is asserted and no aggregation causality is intended, the weaker related-to relationship may be used instead.

5.6.4. The Dialog Graph

The set of Dialogs participating in an Activity can therefore be modeled as a directed graph:

G_D = (D, R_D)

where D is the set of Dialogs and R_D is the set of typed relationships among them. Examples MAY include:

delegated-from
aggregated-from
continuation-of
related-to
replaces

The graph may branch, converge, and evolve over time, and no participant is required to maintain it in full.

5.7. Independent Dialog Lifecycle

Related Dialogs may hold different lifecycle states simultaneously (the states below are illustrative; the dialog-management protocol defines the exhaustive list of lifecycle states of a Dialog.)

D2 ACTIVE
D3 COMPLETED
D5 SUSPENDED

Termination of D5 does not imply termination of D2, and failure of D5 does not necessarily invalidate D2 or D3.

This distinction matters for failure recovery, long-running activities, asynchronous delegation, partial completion, and retry behavior.

Related Dialogs can be correlated (they belong to the same Activity and may be linked through lineage) yet maintain independent lifecycle states. A transition in one Dialog does not inherently impose the same transition on another Dialog or its Tasks. However, changes in requirements, dependencies, or security and authorization conditions may have consequences for related work. Where needed, sufficient Correlation Metadata is preserved and selectively propagated to identify affected Dialogs and Tasks and associate any resulting changes with their causes.

Lifecycle coupling, for example:

if D2 terminates, D5 should terminate

is not implied by this architecture. The actual consequences for related work depend on the applicable protocols, policies, and application requirements. This architecture concerns their correlation, not the operational transitions themselves. A suspended Dialog may be resumed, or a sub-dialog may be created as a child of it to resume partial operations, independently of other Dialogs.

5.8. Recovery and Continuation

A failure of a participant, intermediary, or transport path may interrupt communication without necessarily ending a logical Dialog. Recovery requires distinguishing the continuity of the Dialog's identity from continuity of its work and correlation context.

When the existing Dialog D2 is resumed, its logical identity remains D2, even if its transport connection, protocol session, or protocol-local identifiers change:

Before interruption:  D2
After resumption:     D2     (same logical Dialog)

No continuation-of relationship between distinct Dialogs is needed in this case, because no new logical Dialog has been established. The correlations involving D2 must remain available or be recoverable where they are needed to maintain coherent context.

Alternatively, recovery may involve establishing a new Dialog D2' with a distinct logical identity. Three situations are possible:

  • Continuation without replacement: D2' carries forward work or correlation context associated with D2, but does not necessarily supersede D2.

    D2' continuation-of D2

  • Replacement without continuation: D2' supersedes D2 as a Dialog, for example when the original objective is attempted again without carrying forward D2's previous execution context.

    D2' replaces D2

  • Continuation and replacement: D2' both carries forward work or context associated with D2 and supersedes D2 as a Dialog.

    D2' continuation-of D2
D2' replaces D2

continuation-of expresses continuity of work or correlation context between distinct Dialogs. replaces expresses Dialog-level supersession, rather than the transfer of an arbitrary subset of Tasks. The two relationships are not mutually exclusive. Neither imposes automatic Task lifecycle transitions, termination of the original Dialog, or wholesale inheritance of its correlations. Where only some work is carried forward, continuation-of together with relevant Task or Resource relationships may express that narrower continuity without asserting Dialog-level supersession.

Historical associations remain meaningful after recovery. For example, if D3 was delegated from D2, that fact does not disappear when D2' replaces D2. If D2' also needs to correlate subsequent work with D3, the necessary associations must be preserved or established; they cannot be inferred solely from D2' replaces D2.

Communication recovery and correlation recovery are distinct. A transport may reconnect while a participant has lost the mappings needed to correlate D2 with related Dialogs, Tasks, or Resources. Conversely, those mappings may remain available while a participant or its transport path is unreachable. Resuming the same Dialog identity also does not, by itself, guarantee recovery of its application Task state or of the outcome of an interrupted operation. An uncertain operation outcome is not itself a correlation failure if the operation remains correctly associated with its relevant Dialog and Task.

Within an established protocol-managed correlation scope, a participant remains responsible for the correlations it undertook to maintain through an interruption or recovery. Where necessary correlations cannot be preserved or recovered, that limitation must be distinguishable from successful correlation maintenance, subject to the detection and reporting capabilities of the relevant protocol. This does not imply visibility into unregistered downstream Dialogs or perfect attribution of a recovery failure.

Recovery may involve several distinct layers:

  • Correlation context: preservation or re-establishment of Dialog identities, relationships, and relevant Task and Resource associations, as represented by the context-management protocol.

  • Application-protocol state: the sessions, Tasks, and operation outcomes maintained by protocols such as A2A and MCP, which may continue independently of a transport connection.

  • Transport connectivity: re-establishment or migration of a connection or path through the underlying transport mechanisms.

The architecture identifies these distinct concerns and their correlation implications, without specifying recovery procedures. Open protocol-design questions include how lost mappings are recovered or rebuilt, how peers distinguish resumption from new Dialog creation, how correlation is handled when an intermediary disappears, and how unreachable participants or uncertain operation outcomes are reported. These questions are to be addressed in the dialog-management protocol draft, in coordination with relevant IETF working groups (for example, SIP Core for Dialog semantics, QUIC for transport migration, and TLS, WIMSE, and OAuth for session, identity, and delegated authority) and with the application protocols to which the protocol binds, such as A2A and MCP.

5.9. Interruptions and Barge-in

Work in progress can be interrupted. A barge-in is new input introduced while work is underway that may modify, redirect, or interrupt it. It does not automatically determine whether Tasks are modified, created, or replaced, or whether another Dialog is established.

The running Activity of "analyze the video" shows the pattern. The user asked Agent A to analyze video V1; Agent A delegated the video-analysis Tasks to Agent B and the transcription Task to Tool X, and work is in progress. Ten minutes in, the user barges in and tells Agent A to analyze only the last 50% of the video. In this example, no new entity is invoked: the same Activity and Dialogs remain, while the requirement governing the work changes.

Two aspects of this barge-in are relevant to correlation:

  • First, the new requirement must be associated with the work it affects. In this example, Agent B's Task is modified to analyze only the last 50% of the video. Other barge-ins could instead result in new or replacement Tasks, suspension, cancellation, or no change to particular Tasks. Whether Tool X's transcription Task is also affected is an application decision: the existing transcript may still cover the full video, and Agent B may simply use its relevant part, or Tool X's Task may be narrowed to match.

  • Second, the input that carries the barge-in may arrive in a different modality from the original requirement. The original requirement reached Agent A as text; the barge-in reached it as voice. In this example, the spoken correction is a Resource used as input to the affected Task:

AX1 input-to TB1

The correction is not a standalone requirement: "analyze only the last 50%" is meaningful only relative to the original requirement it narrows. Recording that the spoken correction modifies the original requirement, rather than leaving the two statements unconnected, is what lets any participant determine which requirement governs the work after the interruption. How that association is encoded is left to the dialog-management protocol.

A barge-in may have several correlation-relevant aspects, which are not mutually exclusive:

  • a lifecycle change in an affected Dialog, for example from active to interrupted;

  • the establishment of a related Dialog, such as a clarification side-channel; and

  • Resource provenance or binding information when the new input uses a different modality or changes the requirement associated with a Task.

Which aspects arise is an application decision. The architectural concern is that the initiating input, affected Dialogs, Tasks, and Resources, and any resulting changes can be correctly associated; it does not prescribe the operational response.

A barge-in can also affect in-flight or subsequent Events within an affected Task. Events remain local to their Tasks and do not acquire independent lifecycle states. Where necessary, their association with the affected Task and relevant requirements must remain interpretable.

A barge-in may affect several related Dialogs. Suppose Agent B has created D5 with Agent C, Agent C has created D7, and D8 aggregates D5 and D6. In this example:

  • D5 is interrupted: its in-flight Task is modified, because the requirement it was pursuing has changed.

  • D7, which Agent C chained from D5, is suspended: the work it was doing served the requirement that has just changed.

  • D8, which aggregates results produced independently in D5 and D6, is terminated: the Task that required D8 to aggregate from upstream Dialogs was replaced by the barge-in, so D8's reason for existence is gone and its awaiting Tasks are cancelled.

  • D6, which runs independently of D5, is untouched.

One interruption readjusted work in four Dialogs: a lifecycle transition and Task modification in D5, a suspension in D7, and a termination in D8. Each is an explicit fact of the model, tied back to the interruption through the lineage relationships that link the four Dialogs.

These questions are candidates for coordination with transport- oriented work (for example, WebTransport, WebRTC, and MoQ) and with the dialog-management protocol. Open questions include:

  • how lifecycle transitions behave when interruptions occur mid-course;

  • how changes resulting from a barge-in are associated with affected Dialogs, Tasks, and Resources;

  • how affected ongoing or subsequent Events remain interpretable within the relevant Task;

  • how the association between the barge-in input and the requirement it affects is preserved across modality changes; and

  • how far correlation information about the effects of a barge-in needs to travel through the Activity graph.

5.10. Reuse Across Dialogs

The same Resource can appear in several Dialogs. In the running example, V1 is used by both D2 and D3, and T1, produced in D3, is consumed in D2 and later referenced in D5 as provenance:

Resource V1                 Resource T1
  |                           |
  +--> D2                     +--> D2
  |                           |
  +--> D3                     +--> D5

Sharing a Resource does not imply identical local interpretation:

D2:
  T1 analyzed as primary text input

D5:
  T1 retained only as provenance

Resource identity remains stable while interpretation is local. These Dialog-local interpretations are Resource Bindings, and roles can change as the Activity executes and delegations happen.

Sharing also does not imply the same Dialog. For example, that D2 and D5 both reference V1 does not imply:

D2 == D5

or that Agent A, Agent B, and Agent C share one conversation. Resource correlation and Dialog correlation are separate concerns: a reference to a Resource carries a meaning understandable only within its respective Dialog.

Likewise, a shared Activity does not imply a shared participant set:

D2 belongs-to the Activity
D5 belongs-to the Activity

does not imply that all participants in D2 and D5 know or can address one another. The Activity provides a logical grouping, not a communication group.

5.11. The Running Example, Assembled

The concepts introduced in this section can be assembled into a single picture of the running example. The figure below unifies the two views of the Architectural Model: the invocation view of the interaction wave, instantiated with the example's participants and roles, framed by the Activity root above and the Resource Layer below, as in the containment view.

                   Activity: analyze video V1
                                |
                           User (AIE)
                                |
                                | Dialog D1 (Activity, V1)
                                v
                          Agent A (FNE)
                       + -----------------+
         Dialog D2     |                  |   Dialog D3
      delegated-from D1|                  | delegated-from D1
           task(s)     v                  v  task (one-shot)
                 Agent B (AA)        Tool X (AT)
                       +
      Dialog D5        |
      delegated-from D2|
      task(s)          |
                        v
                  Agent C (AA)
               (invisible to Agent A)

      +----------------------------------------------+
      |               Resource Layer                 |
      |  V1:  transferred to D2; referenced by D3    |
      |  T1:  produced in D3; consumed in D2;        |
      |       referenced as provenance in D5         |
      |  AX1: barge-in input (voice); input-to TB1   |
      +----------------------------------------------+

The figure exhibits, in one place, most of the properties established in this section:

  • Branching: D2 and D3 are both delegated from D1.

  • Chaining: D5 is delegated from D2, extending the wave below Agent B.

  • Hidden delegation: D5 and Agent C are invisible to Agent A.

  • Independent lifecycle: the Dialogs need not share state; D3 completes with its one-shot Task while D2 remains active and D5 may be suspended.

  • Resource reuse: V1 is shared by D2 and D3, and T1, produced in D3, is consumed in D2 and referenced as provenance in D5.

  • Interruption: a barge-in (AX1, voice) narrows the objective of TB1 mid-course; its effects travel through lineage to the Dialogs that depend on it.

Every edge is an invocation, every Dialog is independently maintained, and every correlation is an explicit relationship. This is the structure that context propagation, the subject of the following section, must preserve.

6. Tasks Within a Dialog

Tasks are bounded units of work within Dialogs. An invocation establishes a Dialog but does not necessarily create a Task. A Dialog may contain zero or more Tasks, performed sequentially, in parallel, or through one-shot Assistant Tool exchanges.

The previous section zoomed in from Activities to Dialogs. This section zooms in one more level, to the Tasks inside each Dialog. It examines Task identity, states, cross-Dialog relationships, and the effects of failures and privacy boundaries.

The model defines the relevant semantics; protocol documents determine their implementation.

6.1. Task Identity: Local and Global

Tasks use identifiers defined by the local application protocol (such as an A2A taskId). Such identifiers are not assumed to be meaningful or resolvable outside their Dialog.

Unlike a Resource that needs stable identity across boundaries, a Task need not have a universal identifier. Its cross-Dialog association may be asserted by a participant that can resolve the relevant local Tasks. Whether to standardize a global Task identifier remains a protocol-design decision.

6.2. Task States

Tasks have lifecycle states independent of their Dialog's state. The following are illustrative, not normative; the dialog-management protocol defines the applicable states.

Progression states - a Task holds exactly one at a time:

defined       the Task exists; work has not begun
queued        the Task was accepted and is waiting behind
              other work
in-progress   work is in progress
awaiting      the Task cannot make progress on its own; it is
              blocked on something external, for example another
              Task's result, a decision, or an authorization
delegated     part of the work was passed to another Dialog;
              the Task awaits its result
awaiting-approval   the Task is blocked on a human decision
auth-required       the Task is blocked on additional
                    authorization
interrupted   work was paused or cut off; may resume
completed     the bounded work is done
failed        work ended without completing
timed-out     the Task exceeded its maximum duration
rejected      the work was declined before it began, for
              example out of scope or against policy
cancelled     the Task was ended administratively, for example
              by a revoked delegation
unknown       the state of the Task cannot be determined

Modifiers - facts about the Task that may persist alongside a progression state:

modified      the objective or parameters of the Task changed
superseded    another Task replaces this one

Some states, such as completed, timed-out, rejected, or cancelled, may be final. Retry behavior is an application decision.

A one-shot Task with an Assistant Tool can pass through the whole set in one exchange: defined on assignment, in-progress on submission, completed on response.

Structural relations such as delegation or replacement are generally stable; lifecycle assertions such as completion and modification are mutable and can become stale. Participant acting on an old copy can make incorrect decisions, as discussed later under Stale Context.

A Task has one progression state at a time; modifiers may coexist with it. Thus a Task may be in-progress and modified without changing its progression state.

Lifecycle updates are also Correlation Metadata. "t_32 completed", "t_21 modified", "t_15 superseded" are precisely the assertions that may need to propagate between involved parties, in the upstream direction alongside results and failure information, as discussed later under Directional Propagation.

6.3. Task Independence

Each Task has its own lifecycle. A change to one does not automatically change another, even within the same Dialog.

Within a single Dialog, several Tasks can be in different states at the same time:

t_12 in-progress
t_22 awaiting
t_32 completed

Completing, failing, or cancelling one Task does not automatically change the other Tasks within its Dialog.

Task independence also applies across Dialogs. Relationships such as delegated-from or replaces record association; it does not couple their lifecycles, and it does not cause a change in one Task to propagate to another. Dependent work changes only when its participants decide or applicable requirements demand it.

This permits partial completion, parallel work, and selective failure within one Activity.

As with Dialog lifecycle, a Task's lifecycle state is mutable and may become stale; a participant acting on an old copy can make incorrect decisions, as discussed later under Stale Context.

6.4. Task Relationships Across Dialogs

Examples of Task-level relationship types include:

delegated-from    t_12 delegated-from t_21
related-to        a weaker assertion of correlation, without
                  delegation causality
replaces          a Task supersedes another, typically a failed
                  or abandoned one

delegated-from is the primary cross-Dialog relationship: it asserts that one Task exists because of a requirement found in another. A Dialog may contain many Tasks while only one of them causes a downstream delegation; delegated-from records which one (see "Task Lineage").

At Task granularity, replaces expresses that one Task supersedes another. Replacing a Dialog does not automatically replace all of its Tasks. When a new Task actually supersedes an original Task, its relationship can be expressed explicitly:

t_12' replaces t_12

Dependency on an existing result may be inferred from Resource provenance; awaiting a future result is Task state rather than a new relationship type.

This avoids introducing a separate verb for every application dependency.

6.5. Task Topologies

Dialog and Task lineage are related but distinct. Creating a Dialog does not require creating a Task or Task-to-Task association. Where delegation originates in a Task, its Task relationship can be represented explicitly:

D5 delegated-from D2       (Dialog level)
t_15 delegated-from t_22   (Task level, when applicable)

The Dialog and Task graphs are related but need not mirror each other:

  • A Dialog may contain zero or more Tasks, so Task relationships add finer-grained detail where applicable.

  • Within one Dialog, Tasks relate to one another through their Resources: the output of one Task is the input of another, sequentially or in parallel.

  • One-shot interactions with Assistant Tools form leaves of the task graph: submit, receive, done.

The set of Tasks participating in an Activity can therefore be modeled as a directed graph:

G_T = (T, R_T)

where T contains the Tasks and R_T their typed relationships. This graph may correspond to, but need not mirror, the Dialog graph G_D. No participant MUST maintain it in full.

6.6. Impact of Dialog Failures on Task Relationships

Dialog failures or lifecycle changes may affect Tasks without determining their states. Where consequential changes occur, they need to remain correlated with the affected Dialogs and upstream work.

  • Dialog terminated: Tasks may complete, fail, be cancelled, or continue under application-specific arrangements. Correlation captures relevant consequences without imposing a uniform transition.

  • Dialog replaced: When D2' replaces D2, that Dialog-level assertion does not determine what happens to each Task in D2. A Task may be continued through a new Task, superseded, retained in its existing state, or handled otherwise according to the application. Where a new Task t_12' actually supersedes t_12, that association can be recorded explicitly:

t_12' replaces t_12

A Task that is not re-established in D2' must not automatically be interpreted as lost, cancelled, or failed solely because its Dialog was replaced. Its known outcome or unresolved status, where relevant, remains part of the correlation context.

  • Dialog suspended: Depending on the work and its dependencies, associated Tasks may be interrupted, left awaiting, or continue through other arrangements. The relevant Task states and any resulting changes may need to be correlated with the Dialog's suspension.

  • Broken lineage: Loss of an upstream Dialog relationship can make related cross-Dialog Task assertions unverifiable and prevent result correlation or upstream resolution.

6.7. Privacy Considerations

Task objectives and states can reveal intent, workflow, timing, and delegation. Sharing results need not require disclosing the Tasks that produced them.

Administrative domains may therefore:

  • disclose Dialog participation without Task-level detail;

  • share a Task's output while withholding its objective;

  • omit cross-Dialog task relationships entirely;

  • generalize Task objectives before propagation.

Task associations may be kept privately by intermediaries. Even Task states may be withheld or generalized according to the applicable disclosure policy.

7. Binding to Protocol-Local Context

Application protocols define their own context, Task machinery, and identifiers, and this architecture does not replace them.

The architecture associates Dialogs and Tasks with applicable protocol-local constructs, without redefining their identifiers:

Activity
   |
   +-- Dialog         <-> protocol-local context identifier
   |      |
   |      +-- Task    <-> protocol-local task identifier
   |
   +-- Dialog
          |
          +-- Task

Local protocols retain their identifiers and Task state. The model adds cross-Dialog relationships, such as t_14 delegated-from t_22, through mappings when needed.

A protocol-local identifier is meaningful only within the Dialog in which it was established, and identifiers from different Dialogs or protocols are not comparable. They are therefore not sufficient on their own to relate objects across Dialogs. An assertion such as "t_14 delegated-from t_22" names a Task local to one Dialog and a Task local to another; it can be made only by a participant that can resolve both, and it can reach a further Dialog only when the endpoints are made referenceable there, for example through an opaque reference or a translated identifier. Where correlation must persist across boundaries, the model carries stable logical identities rather than local ones (see "Resource Identity" and "Activity Correlation"). This is the point at which protocol-local context ends and correlation begins.

The boundary is between local identification and cross-Dialog correlation, with shared semantics used where interoperability requires them.

8. Context Management and Correlation Metadata Propagation

The preceding sections defined the model's objects and relationships. This section considers the Correlation Metadata needed to preserve and propagate their associations across participants.

Here, "context" does not mean an AI agent's internal memory, context window, or conversation history.

Agent-internal memory and its content remain outside this architecture's scope.

Here, context means relationships. Context is the set of correlations among Activities, Dialogs, Tasks, and Resources: this video belongs to this Task, this transcript is derived-from that video, this Dialog was delegated-from that one. What travels between parties is Correlation Metadata: small, typed statements about those relationships, such as a Task completed, a Dialog replaced, or a delegation revoked. Context Management therefore means maintaining clear relationships among Activities, Dialogs, Tasks, and Resources. This is achieved by preserving the necessary Correlation Metadata locally and propagating it between participants when required. It does not mean copying agent memory, shipping content, or transferring conversation history.

In this Context Management and correlation metadata propagation model, the responsibility for maintaining coherence is layered:

This distinction frames the propagation problem.

Dialogs branch, converge, or replace one another; Tasks and Resources introduce finer-grained associations. Because their contexts are independently maintained, participants must preserve and selectively propagate Correlation Metadata where necessary.

When extending a Dialog or delegating work, a participant SHOULD decide what to preserve, disclose, translate, or bind locally:

Those choices depend on requirements, privacy and authorization policies, and administrative boundaries. The architecture makes the resulting correlations expressible; the protocol determines how they are communicated.

What a participant discloses is its decision; whether the Activity can proceed on those terms is the consequence. Maintaining correlation among an Activity's Dialogs, Tasks, and Resources requires sufficient Correlation Metadata to establish the necessary associations. However, this does not require every participant to possess the complete correlation context of the Activity. Correlations may be preserved locally by participants that maintain the associations between upstream and downstream Dialogs, Tasks, and Resources.

A participant may withhold Correlation Metadata in accordance with its policies. However, when the information available to the participants responsible for maintaining a required correlation is insufficient, that correlation cannot be established. The affected Dialogs or Tasks may therefore be unable to proceed. The architecture requires that such failures be reported explicitly rather than result in indefinite stalls. Forcing disclosure is a matter of authorization and trust and is outside the scope of this document.

The following subsections identify the relevant scope, metadata, and propagation limits.

8.1. Correlation Scope, Extension, and Maintenance Responsibilities

The logical Activity and its protocol-managed correlation scope may differ. Participants may create downstream Dialogs without enrolling them in an existing correlation protocol. The architecture therefore distinguishes:

Existence of a relationship

from

Relationship preservation by a participant

and from

Relationship inclusion in a protocol-managed correlation scope

An interaction can exist without being included in the protocol's managed scope.

8.1.1. Logical and Managed Correlation Scope

The following are views of one Activity, not additional protocol objects:

  • Logical interaction scope: the Dialogs, Tasks, Resources, and relationships that actually contribute to the Activity, including those unknown to some or all of its participants.

  • Maintained correlation scope: the associations that participants actually preserve and can relate, whether locally or through one or more context-management protocols.

  • Protocol-managed correlation scope: the associations brought under a particular interoperable context-management protocol, with identified participants accepting the corresponding maintenance responsibilities. Such scope need not be globally enumerable or visible to any single participant.

For example, the AIE invokes the FNE through D1, the FNE invokes Agent A through D2, and Agent A invokes Agent B through D3. An illustrative context-management protocol (e.g., MACP) manages their correlations with Act-1. Agent B then invokes Agent C through D5 using another non-interoperable protocol (XYZ), or no context-management protocol:

               Logical interaction scope of Act-1
+-----------------------------------------------------------------+
|                                                                 |
| AIE ---D1--- FNE ---D2--- Agent A ---D3--- Agent B              |
|  |                                           |                  |
|  +--- ABC-managed correlation scope ---------+                  |
|                                              |                  |
|                                              +--- D5 (XYZ)--- C |
|                                                                 |
+-----------------------------------------------------------------+

D5 can be logically related to Act-1 although ABC has no knowledge of D5 or its participants. D5's absence from ABC is not, by itself, a correlation-maintenance failure. In particular, the logical interaction scope under Act-1 may extend beyond the correlation scope of an employed protocol (i.e., ABC or XYZ). In other words, the correlation scope of a protocol follows the relationships that have actually been established for management by that protocol.

In summary, each management protocol has a correlation scope that may span the logical interaction scope of its associated activity. Leakage between correlation scopes is determined by the participants shared across those scopes.

8.1.2. Selective Extension and Boundary Bridging (Scope Leakage)

Extending a protocol-managed correlation to another Dialog is a participant decision. Agent B may still maintain a private bridge between its Task in D3 and work delegated through D5:

           ABC-managed                        Outside ABC
       +------------------+              +-------------------+
       | D3: Task t_13    |              | D5: Task t_15     |
       | Agent A <-> B    |              | Agent B <-> C     |
       +---------+--------+              +---------+---------+
                 |                                 |
                 +-- B's local t_13 <-> t_15 map --+
                              (not disclosed)

B's private mapping can interpret D5 outcomes or relay relevant changes without disclosing C or D5. This preserves local correlation but does not place D5 within the first protocol's managed scope. The architecture does not prescribe protocol adoption or bridging mechanisms.

8.1.3. Maintenance Responsibilities Within the Established Scope

The principle is:

  • Voluntary extension must be distinguished from responsible maintenance.

Once participants establish protocol-managed correlations, they undertake to maintain the associations they accepted. They must preserve or communicate relevant changes to affected correlations, not broadcast every update or disclose the complete upstream context. Intermediaries may retain private mappings or translate identifiers.

Consider a barge-in in D1 changing requirements for Tasks in D2 and D3:

       D1                   D2                       D3
 AIE <----> FNE       FNE <----> agent A   Agent A <----> Agent B
     Task t_11            Task t_22                 Task t_13
        |                      |                       |
 barge-in changes              |                       |
 requirement for t_11          |                       |
        +-- relevant update -->+-- relevant update --->+
                               |                       |
                        correlate impact         correlate impact
                            on t_22                  on t_13

     An unrelated Dialog D5 does not require the same update.

Participants managing D1--D2--D3 have responsibilities for those established associations. If Agent A fails to communicate a necessary change between D2 and D3, that is a managed-correlation maintenance failure, not merely a choice not to extend scope (i.e., a violation). The protocol should make this distinction observable without forcing disclosure of private metadata.

The protocol cannot impose maintenance duties on Agent C in an unenrolled D5. Agent B remains responsible for its managed D3 associations, but failure within private D5 may be unobservable and its root cause uncertain.

8.2. The Propagation Model

Propagation carries Correlation Metadata among participants so separately maintained Dialogs, Tasks, and Resources remain correctly associated with their Activity.

Propagation adapts selected correlation information across Dialog boundaries rather than copying a participant's entire state:

upstream correlation
      |
      | select      (choose the metadata the receiving Dialog needs)
      | translate   (map identifiers across the Dialog boundary)
      | correlate   (relate it to objects already present)
      | bind        (associate it with local objects and state)
      v
downstream correlation

The receiving Dialog's correlation context is related to, but need not duplicate, upstream context. The bridging participant selects the relevant associations.

Metadata may concern Activity association, Dialog or Task lineage, or Resource identity and provenance, as described below.

Two properties follow, both developed later in this section:

  • Propagation is selective: Only the metadata a receiving Dialog needs to maintain continuity, associate related interactions, or interpret shared Resources should be carried.

  • Propagation is directional: What travels downstream differs from what travels upstream.

The architecture does not prescribe a fixed set of propagation events. Propagation is need-driven. A participant selects the Correlation Metadata needed when extending correlation to a new Dialog, subject to applicable policies. Once a participant has accepted responsibility for an established protocol-managed correlation, however, subsequent changes relevant to that association are subject to its maintenance responsibilities rather than a wholly new choice about whether to maintain the correlation at all. Establishing a new Dialog, delegating a Task, transferring a Resource, or resuming an interaction may each create a need for propagation.

Participants may retain more correlations locally than they disclose; receiving Dialogs need only the subset required for coherent interaction.

8.3. What Is Propagated

The main classes of Correlation Metadata are described below.

8.3.1. Activity Correlation

A downstream interaction may need to know that it participates in the same broader Activity as other interactions:

D2 belongs-to Activity A1
D5 belongs-to Activity A1

An Activity-ID is a stable, standardized semantic identity for correlating an Activity's Dialogs, Tasks, and Resources.

Identity and disclosure are separate: a stable Activity-ID may be shared, withheld, aliased, or mapped locally while preserving sufficient correlation where needed.

How an Activity-ID is represented or exchanged is a policy and protocol-design decision (see Identifier Translation).

8.3.2. Dialog Lineage

When a new Dialog results from an existing one, the delegating participant can retain the relationship:

D5 delegated-from D2

Dialog lineage preserves the causal or delegation relationship between two Dialogs. Preserving lineage does not imply propagating or disclosing it. A participant may retain lineage locally, and propagate it only when the receiving Dialog needs it, for example for upstream resolution or result correlation. When propagated, the relationship may be carried directly or as an opaque reference, or it may be withheld. Preservation, propagation, and disclosure are therefore separate concerns.

Delegation is one lineage type among several; continuation-of, replaces, aggregated-from, and related-to are the others, and each propagates on the same terms described here.

8.3.3. Task Lineage

A downstream Task may be correlated with the upstream Task from which it originated:

t_15 delegated-from t_22

Task lineage is useful when a Dialog contains several Tasks and only one of them caused the downstream delegation; delegated-from records which one. Protocol-local Task identifiers are not intended to travel; the relationship carries the correlation. Task lineage is therefore optional and selective: it is established where the receiving Dialog requires it for continuity, upstream resolution, or correct interpretation, and withheld otherwise. Under hidden delegation it may be retained only by the participant common to both Dialogs.

As with Dialog lineage, related-to and replaces are examples of other Task lineage types, and each propagates on the same terms.

8.3.4. Resource Identity

A Resource that is referenced or transferred across Dialog, protocol, or administrative boundaries may need to retain a stable logical identity, so that independently operated participants can recognize that they refer to the same Resource:

Dialog D2: Resource T1
Dialog D5: Resource T1

Where cross-Dialog correlation of a Resource is required, its logical Resource identifier should be preserved across boundaries. Each domain and application protocol remains free to hold its own local representation; the Resource-ID provides the correlation:

Dialog D2: Resource-ID = T1, A2A artifactId = artifact-17
Dialog D5: Resource-ID = T1, A2A artifactId = artifact-91

A Resource's stable logical identity supports cross-Dialog correlation; local identifiers, storage, and representation remain implementation-specific. Content transfer is separate from the correlation model.

Identity identifies the Resource; provenance describes its origins, and Binding describes its local use.

8.3.4.1. Resource Provenance

Resource identity establishes how participants refer to the same Resource. Resource Provenance is another resource-specific typed relationship. Examples include:

T1 derived-from V1
TXT-3 produced as a result of task 1 in dialog 1

TBD

8.3.4.2. Local Resource Binding

The same Resource can play a different role in each Dialog or Task in which it appears. A Resource Binding associates a Resource with a specific Dialog or Task and describes how it is used there:

T1 input-to TB1
T1 augments TC1

Unlike identity and provenance, a Binding is not an end-to-end property of the Resource:

D2 / Task TB1:  T1 primary-input-to TB1
D5 / Task TC1:  T1 supporting-input-to TC1

The same logical Resource is referenced in both cases, but its local role differs. When a Resource is propagated into a new Dialog, the upstream Binding is not assumed to apply there. The delegating party may carry it unchanged, translate it, supply it in part, leave it unspecified, or let the receiving participant establish a new one.

Bindings are therefore local semantics. They may be carried across a boundary when useful, but doing so does not make them universal:

T1 required-for TB1   does not imply   T1 required-for TC1

8.3.5. Propagation Summary

The classes differ in what they are and how they typically travel:

Table 1
Information Role Typical propagation
Activity-ID Stable identity of the Activity When cross-Dialog Activity correlation is needed
Dialog lineage Causal or delegation relation between Dialogs Retained where needed; selectively propagated
Task identifiers Protocol-local unit-of-work tracking Normally local to the application protocol
Resource-ID Stable identity of a logical Resource When cross-Dialog Resource correlation is needed
Resource provenance How a Resource was produced or derived When needed for correct interpretation
Resource Binding Local role of a Resource Primarily local; selectively carried

The rule is constant across all of them:

  • Propagate only the correlation the receiving Dialog needs to maintain continuity, associate related interactions, or interpret shared Resources.

8.4. Propagation Is Not Transitive by Default

Propagation in this architecture is selective rather than automatic.

A contextual element can be stable or important without being transmitted to every participant. Each propagation decision depends on whether the receiving Dialog requires that information for correlation, continuity, or correct interpretation. Receiving context information does not imply that the receiver SHOULD automatically propagate all of it further. For example:

A <----> B <----> C

B may receive from A:

Activity A1
Resource V1
Resource T1
A's identity
private policy metadata

and propagate to C only:

Activity correlation
V1
T1

Therefore:

  • Context received from an upstream participant is not automatically transitive.

This principle governs the selection and disclosure of information when extending correlation. It does not release a participant from maintenance responsibilities already accepted for an established protocol-managed association. A participant need not forward an upstream update unchanged, but must preserve or communicate enough information to maintain the correlations for which it is responsible, or make the resulting inability to maintain them distinguishable.

Each propagation step is subject to a new projection decision. Such decisions are left up to the participating parties to determine as described in section 8.3 context projection. The defined protocol is a tool that allows the selective propagation of context metadata.

This is important when agentic systems cross:

  • organizational boundaries;

  • trust boundaries;

  • privacy boundaries;

  • protocol boundaries; and

  • changes in task extent.

Propagation is triggered by the needs of the receiving Dialog rather than by a fixed set of protocol events. A new Dialog, delegated Task, transferred Resource, resumed interaction, or modality transformation may require propagation, but only the correlation information necessary for that Dialog should be carried.

The architecture therefore defines propagation in terms of information requirements and Dialog boundaries, rather than prescribing a universal set of propagation events.

8.5. Context Projection

Context Projection selects what is disclosed when B creates D2 from D1, without copying D1's complete state. For example:

D1:
  Activity A1
  objective
  V1
  T1
  A's identity
  private policy
  conversation history
  internal reasoning
  credentials

        |
        | projection by B
        v

D2:
  Activity correlation
  delegated task objective
  V1
  T1
  relevant provenance

Projection therefore determines what is exposed to the downstream Dialog. The projection is controlled by:

  • task requirements;

  • privacy policy;

  • authorization policy;

  • administrative-domain policy;

  • protocol capabilities; and

  • application semantics.

The projection algorithm is outside this architecture.

8.6. Identifier Translation

Identifiers may change across domains, provided that the necessary mappings preserve correlation:

Organization 1:
  Activity-ID = A1

Organization 2 external view:
  Activity alias = X7

Organization 2 can retain:

A1 <-> X7

privately.

Likewise, a Resource may be represented as:

Resource T1

in the architecture while having local identifiers:

A2A artifactId = artifact-17
MCP resource URI = ...
local database ID = 9f32

Identifier translation permits correlation without globally visible implementation identifiers. Its mechanisms belong to the protocol bindings or local-protocol specifications.

8.7. Directional Propagation

Propagation can occur in both directions, but the information carried in each direction can differ.

8.7.1. Downstream propagation

Typically carries:

activity correlation
task objective
selected Resources
selected provenance
constraints

8.7.2. Upstream propagation

Typically carries:

results
new Resources
status
questions
failure information
local observations

A downstream participant does not necessarily return the same context that it received.

For example:

B -> C:
  V1
  T1
  analyze video

C -> B:
  result R1
  observation: T1 appears unnecessary

The second message is related to the same Activity but introduces new local information.

8.8. Propagation of Assertions

Correlation Metadata may be expressed as assertions, for example:

D2 delegated-from D1
T1 derived-from V1
T1 input-to TC1

An implementation should distinguish the existence of an assertion from the authority or trustworthiness of the entity making that assertion. For example, receiving:

T1 derived-from V1

does not cryptographically prove the derivation. Likewise:

D2 delegated-from D1

does not by itself prove that the delegation was authorized.

Authentication, authorization, integrity protection, attestation, and provenance mechanisms can be used independently where required. This should be coordinated between the agentproto working group and other respective IETF working groups such as WIMSE and OAuth.

Correlation assertions may also come from different sources. A Dialog's or a Task's state may be reported by its participants, observed by a peer, or asserted by a supervisory system. When two sources disagree about the same object - for example, when one reports a Task as in-progress while another reports it as failed - correlation requires that the assertions remain associated with the correct object and, where needed, distinguishable by their sources and relevant context. This architecture does not determine which assertion is authoritative or prescribe how disagreements are resolved.

8.8.1. Lifecycle-Affecting Requests and Consequences

A lifecycle-affecting request, such as a request to cancel a Task, is distinct from an assertion that the requested transition has occurred. Application protocols may carry the request and report its outcome. This architecture concerns the Correlation Metadata needed to associate such requests and assertions with the affected Dialog or Task, and to maintain correlations with dependent work in other Dialogs.

When an interruption, barge-in, cancellation, or security or authorization change may affect related work, the necessary associations may be preserved by a participant with visibility into the relevant Dialogs and Tasks. Sufficient Correlation Metadata is propagated where needed to allow affected participants to interpret the change and associate resulting states or outcomes with it. The metadata need not expose the full upstream context. Security- or authorization-related information may be included where required for correlation; this architecture does not define authorization decisions or enforcement.

8.9. Context Freshness and Lifetime

Not all correlation assertions have the same lifetime. Examples include:

Dialog D2 terminated
delegation revoked
resource superseded
task completed

Correlation information may be:

stable
mutable
time-bounded
revocable

For example:

T1 derived-from V1

is typically stable provenance.

By contrast:

D2 active

is mutable lifecycle state.

Propagation mechanisms need appropriate freshness and update behavior for each class.

8.10. Propagation Revocation

Revocation invalidates a relationship as a basis for future propagation. For example:

D2 delegated-from D1

is revoked, no new downstream propagation should be created based on that relationship.

Revocation need not terminate the Dialog or erase the historical fact that the relationship existed.

Protocol-design questions include whether revocation cascades, how history is treated, and what participants do with earlier correlations derived from a revoked relationship.

8.11. Partial Visibility

No participant needs the complete Activity graph. For example:

For example:

            A
            |
           D1
            |
            B
           / \
         D2   Tool X
         |
         C

A may know only:

D1
B

C may know only:

D2
B
V1
T1

B may know the relationship between D1 and D2.

Tool X may know only:

V1
its local invocation
T1

The Activity can remain coherent with partial visibility. An intermediary may conceal a downstream Dialog while preserving necessary correlation, or establish it outside managed scope. Only accepted managed correlations entail the corresponding maintenance obligations.

8.12. Propagation Failure Modes

Correlation can fail even when application Tasks and transport operate correctly. The following failures are distinct from application or connectivity failures.

Propagation failures merit separate treatment.

Illustrative failure modes follow. These are just examples. The list may be extended as the work progresses

8.12.1. Loss of Correlation

A participant may receive an interaction or Resource without being able to associate it with the correct Activity, Dialog, Task, or Resource.

For example:

Agent D receives T1

but cannot determine which ongoing Activity or Dialog T1 belongs to.

The Resource may remain usable while cross-Dialog continuity is lost.

Loss of correlation can result from:

  • Missing Activity correlation;

  • Loss of Dialog lineage at an intermediary;

  • Failure to preserve Resource identity; or

  • Inability to map protocol-local identifiers to the broader correlation model.

8.12.2. Incorrect Correlation

Incorrect correlation associates unrelated interactions or Resources. For example:

For example:

T1 belongs-to Activity A1

may be incorrectly interpreted as:

T1 belongs-to Activity A2

Missed bindings may combine unrelated Activities or deliver results to the wrong Dialog; correlation mappings are integrity-sensitive.

Implementations SHOULD therefore treat correlation identifiers and relationship mappings as integrity-sensitive state.

8.12.3. Over-Propagation

A participant may receive more metadata than needed. For example, a summarizer requiring only T1 might receive:

For example, Agent D may require only T1 for summarization but receive:

Activity-ID A1
D3 delegated-from D2
D2 delegated-from D1
T1 derived-from V1
identities of upstream participants

even though none of that information is necessary for summarization.

Over-propagation increases:

  • privacy exposure;

  • topology disclosure;

  • metadata size and overhead transaction;

  • implementation coupling; and

  • cross-domain tracking risk.

Selective propagation protects privacy and reduces coupling.

8.12.4. Stale Context

Some context changes over time:

For example:

D2 ACTIVE

may later become:

D2 TERMINATED

or a propagation relationship may be revoked.

Stale context can produce incorrect decisions. Stable relationships and mutable lifecycle assertions require different freshness handling.

For example, stable provenance differs from mutable Dialog state:

T1 derived-from V1

from mutable state, such as:

D2 ACTIVE

and allow propagation mechanisms to define appropriate freshness and update behavior.

8.12.5. Semantic Drift

A downstream Dialog may incorrectly reuse an upstream local interpretation. For example:

For example:

D1 / TB1:
  T1 required-for TB1

must not automatically be interpreted as:

D2 / TC1:
  T1 required-for TC1

The Resource identity is preserved, but its local Binding may differ.

Semantic drift arises when a local Resource Binding is treated as universally applicable.

8.12.6. Broken Lineage

An intermediary may lose the relation connecting a delegated Dialog to its origin. For example:

For example:

D1 -> D2

is created, but B later loses the relationship:

D2 delegated-from D1

Without that link, subsequent questions, results, and failures may not be routed or correlated upstream.

Broken lineage obstructs:

  • upstream resolution;

  • result correlation;

  • failure recovery; and

  • audit or troubleshooting of delegated execution.

8.12.7. Excessive Lineage Propagation

Excessive lineage disclosure is the opposite risk:

A system may preserve and propagate an unnecessarily complete delegation chain:

D5 <- D4 <- D3 <- D2 <- D1

to every downstream participant.

An accurate but over-shared chain may reveal unnecessary participants, domains, or workflow topology.

Preservation and disclosure are distinct.

8.12.8. Inconsistent Local Mapping

Local mappings may include:

Activity A1 <-> local activity-773

Resource T1 <-> local artifact-19

Inconsistent mappings can bind a valid logical identifier to the wrong local object, appearing externally as a cross-domain correlation failure.

Such failures are local implementation failures, but they can appear externally as incorrect cross-domain correlation.

8.12.9. Partial Propagation Failure

In a multi-element propagation operation, some context may arrive while other required context is lost.

For example, a participant may receive:

Resource T1
Activity-ID A1

but not the Resource identifier or relationship necessary to associate T1 with the Task awaiting it.

Partial propagation can be more difficult to detect than complete propagation failure because the receiving participant may have enough information to continue execution while silently losing intended correlation.

8.12.10. Architectural Implication

These failure modes reinforce a central principle of this architecture:

  • Correct correlation requires preserving sufficient information to maintain the relationships among an Activity's Dialogs, Tasks, and Resources. Such information may be maintained locally or selectively propagated between participants, depending on which associations need to be established and by whom.

A propagation mechanism SHOULD therefore support detection or mitigation of missing, stale, inconsistent, or excessive correlation information without requiring every participant to maintain the complete Activity graph. In a protocol-managed scope, the architecture further recognizes the need to distinguish failures to maintain accepted correlations from interactions that were never included in that scope. Where supported by evidence, the affected managed step and its responsible participant should be identifiable. This does not imply complete visibility into undisclosed downstream interactions or perfect attribution of the underlying cause of failure.

8.13. Propagation Example

The video-analysis example can be expressed as follows.

Initial Dialog:

Activity A1

D1:
  A <-> B

Task TB1:
  V1 input-to TB1

Tool transformation:

A -> Tool X

T1 derived-from V1
T1 output-of IX1

T1 enters D1:

T1 input-to TB1

B creates a downstream Dialog:

D2 delegated-from D1
TC1 delegated-from TB1

B projects:

Activity correlation
V1
T1
T1 derived-from V1
task objective

into D2.

C develops local context:

T1 input-to TC1
T1 appears unnecessary

C asks B for clarification.

B uses locally retained lineage:

D2 delegated-from D1
TC1 delegated-from TB1

to resolve the question upstream.

The complete upstream Dialog does not need to be copied into D2, and the downstream Dialog does not need to be exposed to A.

This is the intended meaning of context propagation in this architecture.

8.14. Relationship Taxonomy

The following are examples of relationship types that are useful at the architecture level. They are illustrative, not an exhaustive or normative set; the dialog-management protocol defines the relationship types that are interoperable.

TBD

8.16. Relationship to Dialog-Management Protocols

A dialog-management protocol may standardize only the subset of the architectural model needed for interoperability. For example:

For example, a protocol may define:

Dialog-ID
Dialog-Relationship
Lifecycle State
Propagation State

without defining:

required-for
optional-for
application-specific planning semantics

This leaves application-specific semantics outside the interoperable baseline.

8.16.1. Architectural vs. Protocol Surface

The architecture defines Activities, Dialogs, Tasks, Resources, and typed Relationships; it does not require every concept to become a standardized protocol field.

The agentproto charter identifies Dialog identifiers, lifecycle semantics, and Dialog-context propagation as subjects for its standards-track protocol work. This information model examines related correlation concepts that may inform that work, including:

  • Dialog identity and identifier aliases;

  • Dialog lineage and other typed relationships (e.g., delegated-from, aggregated-from, continuation-of, related-to, replaces);

  • Dialog lifecycle and propagation assertions;

  • Activity and Task associations where cross-Dialog correlation needs them; and

  • Resource identity or provenance associations relevant to correlating interactions across modalities or tool invocations.

Protocol design determines which information is exchanged, locally retained, or supported by application protocols. Resource Binding remains primarily an application interpretation.

8.17. Relationship to A2A

A2A provides useful local constructs such as:

contextId
taskId
artifactId
messageId

These map naturally into the architecture:

A2A contextId -> protocol-local Dialog context

A2A taskId -> Task

A2A artifactId -> local representation of a Resource

This architecture does not replace those constructs.

It provides a model for correlating them with objects outside one A2A context.

8.21. Architectural Requirements

The following capabilities inform protocol design; they do not each require distinct wire fields:

  1. multiple Dialogs within one Activity;

  2. independent protocol-local context identifiers;

  3. stable Resource correlation where useful;

  4. delegation lineage;

  5. selective projection;

  6. hidden downstream delegation;

  7. cross-domain operation;

  8. identifier translation;

  9. provenance correlation;

  10. privacy-preserving correlation;

  11. correlation of lifecycle-affecting requests, resulting assertions, and dependency-driven consequences across Dialogs and Tasks;

  12. distinction between the logical Activity scope and correlations established or maintained under a particular protocol;

  13. selective extension of managed correlation, with responsibilities for maintaining the associations participants have accepted; and

  14. detection or reporting of failures within managed correlation scope, with bounded identification of affected steps and responsible participants where supported by available evidence.

8.22. Out-of-Scope

This document does not define:

  • An agent's internal memory;

  • Prompts or retrieved documents;

  • Chain of thought;

  • Agent planning semantics;

  • Authorization decisions or enforcement mechanisms;

  • Delegated-authority mechanisms;

  • Trust establishment mechanisms;

  • Agent discovery mechanisms;

  • A wire format.

8.23. Open Questions

  1. Activity and Task are architectural concepts, not baseline protocol objects. The exact protocol surface for optional Activity and task-lineage extensions is to be determined in the protocol draft.

  2. Should Dialog lineage be visible downstream or remain local by default?

  3. What is the minimum Resource correlation required for multimodal dialog continuity?

  4. Should Resource Binding semantics remain entirely application-defined?

  5. How should logical Activity identity survive identifier translation?

  6. How should partial resources such as video segments be represented?

  7. Which relationships require integrity protection?

  8. How should context-resolution questions be correlated across Dialogs?

  9. How much lineage should cross administrative boundaries?

  10. Which subset of this model belongs in a dialog-management protocol?

8.24. IANA Considerations

This document requests no IANA actions.

8.25. Security Considerations

Correlation metadata can reveal workflows, delegation, and Resource relationships.

Implementations should minimize propagated information, avoid semantic identifiers, and treat correlation identifiers as non-authoritative.

Possession of an identifier does not imply authorization.

Relationship assertions may require separate integrity or provenance mechanisms.

8.26. Privacy Considerations

A distributed Activity MUST NOT require a globally visible participant list or a globally stable identifier.

Implementations should:

  • minimize propagated metadata;

  • preserve participant opacity where possible;

  • translate identifiers at administrative boundaries;

  • avoid propagating private dialog history;

  • limit retention of correlation mappings.

8.27. Appendix A. Canonical Architecture 

Activity
 |
 +-- Dialog
 |    |
 |    +-- Task
 |         |
 |         +-- Resource
 |
 +-- Dialog
      |
      +-- Task
           |
           +-- Resource

Illustrative relationships:

Dialog belongs-to Activity
Task belongs-to Dialog
Dialog delegated-from Dialog
Task delegated-from Task
Resource derived-from Resource
Resource input-to Task
Resource output-of Task

Contributors

Hesham Moussa
Huawei Canada
Arashmid Akhavain
Huawei Canada

Authors' Addresses

Hesham Moussa
Huawei Canada
Arashmid Akhavain
Huawei Canada