AAP – An open standard for verifiable AI agent actions (RFC draft)

2 points

3 months ago

As agents move from chat to action — invoking tools, delegating to sub-agents, modifying real systems — there's no standard for what a verifiable agent action looks like. Every framework logs differently. Nothing is interoperable. Nothing is cryptographically verifiable. And nobody has defined what "an agent action" even is, formally. We wrote a minimal RFC to fix that: the Agent Action Protocol (AAP). AAP defines:

A canonical AgentAction data model A cryptographic chain integrity model (SHA256-linked, like a git commit graph) Explicit privacy boundaries — raw reasoning and parameters must never be stored, only their hashes An extensibility mechanism for LangChain, OpenAI Agents, Anthropic, etc.

AAP is explicitly NOT a policy engine, a certification body, or a SaaS. It's a primitive. The smallest possible unit of verifiable agent cognition. One design decision we want feedback on: we made abort a first-class decision type, equal to invoke_tool or delegate. An agent refusing to act is as auditable as an agent acting. Does this hold up in your mental model? Reference implementation (Python) + RFC: https://github.com/Thinklanceai/aap-python Tear it apart.