An agent drafts a reply to an insurance claim. A human reads it, changes one sentence that would have committed the insurer to something it did not mean, and ships it. That edit is the single most valuable thing that happened in the whole interaction. It is a human catching an error a model could not.
Now ask where that edit is recorded. In most production systems, the answer is: in application code, a chat thread, a ticket comment, or nobody’s memory at all. The judgement that mattered most leaves no durable trace. Six months later, when a regulator or a customer asks who decided what and why, you cannot answer.
This is not a niche problem, and it is getting bigger. Foundation models are moving out of “generate a response” and into operational roles: they plan across steps, call tools, request human input, and coordinate with other agents on work that touches customers, claims, code, contracts and clinical decisions. Real deployments are no longer one human supervising one model. They are multi-human, multi-agent collaborations that cross teams, time zones and trust boundaries. And with the EU AI Act’s oversight obligations arriving, “a human was in command, and here is the record” is becoming something you have to be able to show, not just assert.
Two protocols already standardise parts of this world. MCP standardised how an agent reaches tools and data. A2A standardised how agents talk to each other. Neither defines the shared workspace in which humans and agents do accountable work together. That gap is what CHAP fills.
CHAP, the Collaborative Human-Agent Protocol, treats human judgement as a first-class, recorded event. Under CHAP:
- An override becomes a structured event that carries the diff, the rationale, and a content hash. The change, the reason for it, and proof it was not altered later.
- A human approval becomes a signed, non-repudiable decision that can be replayed years later.
- A handoff between shifts becomes a portable envelope, not a pinned message that the next person may never read.
- All of it lands in an append-only evidence log.
The design is deliberately small at the centre and extensible at the edges. A minimal Core defines workspaces, participants, tasks, artefacts, and the evidence log. Composable profiles add review, routing, deliberation, handoff, identity, signatures, and transparency-backed audit, so a light internal tool and a heavily regulated clinical system can both use the same protocol and turn on only what they need.
Everything is open. The specification is published under CC-BY 4.0 and the code under Apache 2.0, with two interoperable reference implementations and adapters for the main agent frameworks (Pydantic AI, LlamaIndex, AG2 and Google ADK), so you can add it to an agent you already have.
It is early, and the point of publishing it now is to be argued with. If you build agents that carry real responsibility, we would like you to read the model, try it against your own workflow, and tell us where it does not hold up.
답글 남기기