[Agentproto] Re: Use case input: cost attribution across agentic dialogs

[Girish Konda <[e-mail]>]

Hey Mehek,

A couple of comments on the cost attribution case and the dialog identifier
is the right primitive for it.

The correlation half holds and the reporting half is where I'd add
something. A participant that can't establish its own usage for a dialog is
not in the same position as one that spent nothing, and if both report zero
the operator can't separate a genuinely free hop from a meter that was
unreachable. That is how cost goes missing rather than double counted, and
a retry on the same dialog then bills a second time against the first
identifier.

I filed draft-konda-agentproto-evaluation-state-00 on this shape of
problem,
https://datatracker.ietf.org/doc/draft-konda-agentproto-evaluation-state/.
Section 4.4 is the stale consumption source case and it already names
consumption state as an input a decision point has to keep separable.

Please provide the write up for the use case and I'm happy to add it to -01
of the draft and acknowledge you for it.

Thanks,
Girish


On Sun, Oct 4, 2026 at 9:36 PM Mehek Mandal <[e-mail]> wrote:

> Hello agentproto,
>
> I'm Mehek Mandal, founder of Traeco (
> https://www.google.com/url?q=http://traeco.dev&source=gmail&ust=1791261249718000&sa=E)
> where we build cost observability for AI agents. Following the thread on
> existing practice, use cases, and deployment models, I'd like to offer a
> deployment use case that depends on dialog correlation.
>
> When a user task fans out across several agents and tools, operators need
> to attribute model and tool spend back to the originating task and to each
> participant. Today that is done with vendor-specific trace IDs that break
> at agent boundaries and intermediaries, so cost is lost or double counted.
> A stable dialog identifier, plus parent and child relationships between
> delegated sub-dialogs, propagated through the dialog context, would let
> each participant report usage against the same dialog without exchanging
> any model content. That seems to sit inside the proposed charter's focus on
> identifiers, lifecycle, and context propagation, and outside the content
> exclusion.
>
> I'm happy to write this up in more detail for the use case or requirements
> document if the group finds it useful.
>
> Best,
> Mehek Mandal
> _______________________________________________
> Agentproto mailing list -- [e-mail]
> To unsubscribe send an email to [e-mail]
>