GitHub - bodaro/a2a-agent-profiles: Two companion A2A extension specifications: the Agent Transparency Profile (signed per-answer disclosures) and the Agent Dialog Profile (binding verification dialog between agents) · GitHub
Skip to content

About

Two companion A2A extension specifications: the Agent Transparency Profile (signed per-answer disclosures) and the Agent Dialog Profile (binding verification dialog between agents)

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

The Agent Transparency Profile & The Agent Dialog Profile

Two companion extension specifications for the A2A (Agent2Agent) protocol, addressing the same gap from two sides: an agent that delegates a goal to another agent receives an answer it has no standard way to appraise.

  • The Agent Transparency Profile (ATP) standardizes the disclosures that accompany a delegated answer: twelve fields in four categories — epistemic (confidence with a testable meaning, evidence, limitations), provenance (actions taken, sources examined, provider revision), operational (cost, elapsed time), and safety attestations — each field tagged by its reporter (recorded by the provider's infrastructure, self-reported by its model, or asserted but mechanically checkable), the whole block signed by the provider. Every field exists because one of the four decisions a receiving agent faces — trust, audit, budget, admit — uses it.

  • The Agent Dialog Profile (ADP) standardizes questioning the answer: a small, open set of typed dialog turns that let one agent require another to certify what it says — to declare, in a signed and receipted turn, that a statement is stood behind rather than offered as musing. Six turn types, four of which carry force and are signed; a hash-chained commitment store derivable from the transcript both parties hold; contradictions adjudicated through the dialog itself.

The pairing in one line: ATP is the testimony a provider files with its answer; ADP is cross-examining the witness.

Both profiles assume dishonest counterparties and say plainly what they do and do not guarantee: they do not make statements true — they make refusal visible, evasion recorded, honesty about error cheaper than concealment, and exposure, when it comes, attributable.

Status

Version 0.9 public drafts (2026-08-17), seeking review. Both papers are complete arguments with worked examples, adversarial analyses, and conformance ladders; the normative JSON Schemas and the reference implementation are in progress (see ROADMAP.md).

Reading order

Each paper stands alone, but ATP first is the natural order: ADP's dialog probes the answers ATP's disclosures describe, and the two share one worked example — a codebase-expert agent answering, then defending, a change-safety question.

Feedback

Issues and pull requests are welcome, on substance or prose. The specifications follow a strict admission rule — no field and no turn type without a named decision or consequence that uses it — and proposals are easiest to evaluate when framed that way.

Author

Mike Papper

License

The specification documents are licensed under CC BY 4.0. Planned reference-implementation code will be licensed under Apache-2.0.

About

Two companion A2A extension specifications: the Agent Transparency Profile (signed per-answer disclosures) and the Agent Dialog Profile (binding verification dialog between agents)

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors