| Internet-Draft | Presented Output Use Case | November 2026 |
| Agrawal | Expires 19 May 2027 | [Page] |
When an agent's output is presented to a user over time, as with synthesized speech, the agent has usually generated more than the user has received at the moment an exchange stops. Existing requirements cover abandoning an exchange and cancelling a task. They do not cover telling the participants in the dialog how much of the output the user actually received. This document describes one use case with two examples, a travel booking and a voice purchase, and derives five requirements that can be incorporated into the working group's use cases and requirements document.¶
This note is to be removed before publishing as an RFC.¶
Discussion of this document takes place on the agentproto Working Group mailing list (mailto:agentproto@ietf.org), which is archived at https://mailarchive.ietf.org/arch/browse/agentproto/. The text was first sent to that list under the subject "Use case input: reporting how much output the user actually received" and revised there.¶
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 19 May 2027.¶
Copyright (c) 2026 IETF Trust and the persons identified as the document authors. All rights reserved.¶
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License.¶
The agentic AI use cases and requirements document [I-D.agentic-ai-usecases-requirements] requires incremental streaming of agent output (A1-1) and a task cancellation message (A1-2). The contact center use cases document [I-D.howe-agentproto-contact-center-usecases] adds identifiers for individual exchanges (CC4-2), a signal that an exchange in progress is abandoned, such as when a user interrupts synthesized speech (CC4-3), and receive-only participants such as a recorder (CC4-4).¶
These requirements describe what was generated and when an exchange stopped. None of them tells the participants how much of the output had been presented to the user when it stopped. For text shown at once the difference is small. For speech it is not: presentation takes longer than generation, so an agent can have produced several sentences of which the user heard half of one.¶
The gap matters to three kinds of participant. The agent that generated the output cannot resolve the user's next utterance if it does not know what the user heard. A recorder holds words the user never received. An agent that the first agent delegated to may treat information as disclosed to the user when it was not.¶
This document proposes one use case and five requirements to close that gap. It is written as input to [I-D.agentic-ai-usecases-requirements] and follows that document's structure for a use case. The requirement identifiers used here (UP-1 to UP-5) are provisional and are for the editors to assign.¶
This document uses the terms Dialog, Dialog Context, Task Context, Task, Modality, Delegation and User as defined in [I-D.agentic-ai-usecases-requirements], and the term exchange as used in requirement CC4-2 of [I-D.howe-agentproto-contact-center-usecases]: an individual unit within a dialog, such as a single utterance or turn. In addition:¶
This document uses "is required to" in the way [I-D.agentic-ai-usecases-requirements] does, to state a requirement on the protocol to be designed. It does not use the key words of BCP 14.¶
An agent streams a spoken reply to a user through a presenting participant. Presentation takes longer than generation, so when the exchange stops, the agent has generated more than the user received. Two events cause this: the user interrupts, or the device does, for example when an alarm or a doorbell takes the audio output.¶
The presenting participant is the only participant that knows where presentation stopped. In a simple deployment it is the client application. In the voice use case of [I-D.howe-agentproto-contact-center-usecases] it is a synthesizer that may be a separate service run by a different provider from the agent, with a recorder and other receive-only participants in the same dialog.¶
Steps 4 and 5 are the ones that existing requirements do not cover.¶
The agent, a recorder (CC4-4) and any agent it delegated to still hold the full text. Without more information:¶
This example was contributed by Francesco Marinoni Moretto.¶
The merchant's agent and a recorder (CC4-4) still hold the full statement. Without more information:¶
Where a seller needs to show what a consumer was told before agreeing, the record has to reflect what was presented, not what was generated. That is the reason the recorder needs the presentation point here. Consumer protection rules of this kind exist in several jurisdictions; [EU-CRD] and [EU-FIC] are examples.¶
The requirements below use the tags of [I-D.agentic-ai-usecases-requirements]. Each is followed by a note on why the existing requirements do not already cover it.¶
The protocol is required to allow the participant that presents output to a user to report, for an exchange, the point up to which that output was presented, referencing the exchange identifier (CC4-2).¶
Not covered today: A1-1 covers delivery of partial output to the client, and CC4-2 identifies an exchange as a whole. Neither lets a participant say how much of the exchange reached the user.¶
When an exchange is abandoned (CC4-3), the report in UP-1 is required to be carried with, or immediately after, the abandon signal, without delaying that signal. This complements A1-1 and A1-2.¶
Not covered today: CC4-3 bounds the latency of the abandon signal and A1-2 defines task cancellation. Both say that output stopped. Neither carries how much had been presented, nor relates that information to the signal in time.¶
The presentation point is required to be expressed in terms that the participant that generated the output can map to its own output, without knowledge of the presenting participant's internals. The representation of the point is left to the application.¶
Not covered today: no existing requirement gives the generating and presenting participants a common reference to a point inside an exchange. Without one, a report from a synthesizer run by another provider cannot be interpreted by the agent.¶
The report is required to reach the participant that generated the output and the other authorized participants in the dialog, including receive-only participants (CC4-4), so that dialog context reflects what was presented rather than what was generated. This applies the dialog context preservation of A1-10 to what was presented.¶
Not covered today: CC4-4 delivers authorized dialog context and lifecycle events to receive-only participants, and A1-10 and CC4-1 preserve dialog context across a change of modality. The presentation point is not among the things they deliver, so a recorder or a delegated agent learns that an exchange was abandoned but not what the user had received.¶
The report is required to distinguish a known presentation point from one that could not be determined, in the way B3-10 distinguishes a cancellation that took effect from one that could not be determined. Whether presented content counts as understood by the user is left to the application.¶
Not covered today: B3-10 reports the outcome of a cancellation along a delegation chain. It says nothing about presentation to a user. Without UP-5, a missing report is indistinguishable from complete presentation.¶
The use of UP-1 after an interruption also supports resumption (B1-6, CC3-1).¶
The presentation point could be left to each application. The list discussion favoured the dialog layer, for this reason: the recipients named in UP-4, recorders and delegated agents, are the participants least likely to see application-layer messages. If the report were only an application-layer message, the receive-only participants would be the first to miss it. What the point means for the user stays with the application, as UP-5 says.¶
[I-D.agentic-ai-usecases-requirements] defines dialog context as the identifiers and lifecycle state needed to correlate and maintain a dialog, and excludes the content exchanged with AI models. The report in UP-1 is consistent with that definition. It is a reference to a point in an identified exchange, together with whether that point is known. It carries no content. A participant that already holds the content of the exchange can apply the point to it; a participant that does not hold the content learns nothing about it from the report.¶
In the discussion of the use cases document, one of its editors has raised reporting the outcome of a cancellation as a question for the working group. This use case is the user-facing part of that question. B3-10 covers whether a cancellation took effect. UP-1, UP-2 and UP-5 cover how much of the output had been presented to the user when it did. Since the report refers to an exchange, it applies whether a dialog carries one task or several.¶
[I-D.liu-moq-live-agent-interaction] defines a barge-in signal from the client and an acknowledgement in which the agent gives the position where its output stopped. That is the generating side's view of where generation ended. The requirements here concern the opposite direction: the presenting side's view of where presentation to the user ended, and its delivery to the other participants in the dialog. The two are complementary.¶
[I-D.agentic-ai-usecases-requirements] places accountability and auditing out of scope. This document does not define an audit record. It asks only that the dialog carry the presentation point, so that any record a participant keeps can be accurate about what the user received.¶
A presentation point describes user behaviour: it reveals when a user interrupted and how much the user listened to. It is to be delivered only to participants authorized to receive dialog context for that dialog, as UP-4 states, and it inherits the considerations for passive participants in [I-D.howe-agentproto-contact-center-usecases].¶
A report may be relied on later to show what a user was told. A presenting participant that reports a later point than the one actually reached makes a record claim more than the user received. The report therefore needs integrity protection and needs to be attributable to the participant that made it, using the protocol's mechanisms for authenticating participants and protecting dialog context. A participant that receives a report can check only that the point lies within the output of the exchange; it cannot check that the user perceived the output, which is why UP-5 leaves that question to the application.¶
UP-2 requires that the report not delay the abandon signal. An implementation that waited for an exact presentation point before signalling abandonment would weaken the bounded latency that CC4-3 asks for.¶
This document has no IANA actions.¶
Francesco Marinoni Moretto contributed the voice purchase example, the consumer protection references and the argument for placing the report in the dialog layer. Konrad Gruszka pointed out the link to authorization checkpoint responses (B2-5). Both reviewed the text on the agentproto mailing list.¶