MARC-Carrying over A2A — interoperability experiment

[c4tzzz]

Hi,

I’m working on MARC, an IETF Internet-Draft for structured control and uncertainty metadata for generative models and agents.

I’m looking at whether MARC-Carrying could be mapped cleanly over A2A, especially around Tasks, Messages, or Artifacts.

Draft:
https://datatracker.ietf.org/doc/draft-c4tz-marc/

Implementation:
https://github.com/c4tzzz/MARC

I’m looking for feedback and, ideally, someone interested in trying an independent A2A integration or interoperability experiment.

Thanks,
c4tz

[msampathkumar]

Hi @c4tzzz If you like to join and present your idea to TSC team, then you can https://github.com/a2aproject/A2A/blob/main/GOVERNANCE.md; Do you wish to provide an A2A extension?

Also I did a quick read of the this proposal and I can share my initial review. Control and Uncertainty Disclosure are definitely some good points.

Here are a few reason I am not convinced its useful in A2A

1. In A2A, one the of the key-features is Secure & Opaque. So an Agent's internals include metadata or uncertainty fall on the Agent itself. A2A's focus is in the agent to agent communication & connectivity.

2. For me scoring ambiguity or uncertainty, is like asking is it hot or cold ? It is a relative concept to me.

[capitaharlock]

For MARC-Carrying over A2A, could you share one small lifecycle example showing which metadata belongs to the Task, Message, or Artifact at each step? That would let an independent integration test whether those associations survive, not just whether the metadata arrives.

[c4tzzz]

Thanks @capitaharlock and @msampathkumar.

I added a small MARC-over-A2A lifecycle example and an interop fixture:

https://github.com/c4tzzz/MARC/blob/main/docs/a2a-lifecycle-example.md

The mapping is:

• no MARC metadata on the Task;
• CLARIFY / ambiguity on the clarification Message;
• RETRIEVE / missing_evidence on the retrieval Message;
• ANSWER / high / answer on the final Artifact.

Only MARC-Disclosure is carried. MARC-Core and the numeric scores stay local to the agent.

Fixture:
https://github.com/c4tzzz/MARC/blob/main/interoperability/a2a-marc-disclosure-lifecycle.json

This should make it possible to test that the metadata stays attached to the right A2A object across the lifecycle.

I’d be interested in testing this as an optional experimental profile using A2A’s existing extension points, if it makes sense on your side. It does not modify A2A core objects or expose MARC-Core internals.

I’m also happy to present it to the TSC. Let me know how I can get it on a future agenda.

[c4tzzz]

Hi @msampathkumar,
Following up on your earlier question about an A2A extension: I already added a lifecycle mapping and interop fixture here:
https://github.com/c4tzzz/MARC/blob/main/interoperability/a2a-marc-disclosure-lifecycle.json
It’s not a formal A2A extension spec yet. If you think that would be useful, I can turn it into one.
Thanks,
c4tz

[kaspermuutra-dotcom]

One angle on the "is this useful in A2A" question: uncertainty metadata becomes much more useful to a calling agent if it can be checked. A declared confidence or control signal on an Artifact is still a self-report. If MARC fields were stable and machine-readable, an external tester could compare declared uncertainty against observed correctness over repeated calls, which would turn MARC from "the agent says it's unsure" into "the agent's stated uncertainty is calibrated". That might be a stronger argument for carrying it than disclosure alone.

Kasper

[c4tzzz]

Thanks Kasper, good point.

MARC would provide a consistent format, while a separate evaluator could compare the declared confidence with the actual results over multiple calls.

I can add a small test fixture for this next to the A2A lifecycle example. Would you mind taking a quick look at the format?

Thanks,
c4tz