INCLUDING THE RESPONSIBLE HUMAN

The agent, or ‘agentic’, era already has two key protocols:

1. MCP: Model Context Protocol - connects agents to tools; and

2. A2A: Agent-to-agent protocol - connects agents to each other.

Both are good - and both are being adopted to excellent effect. But both leave out one critical party a regulated business can't treat as optional: the human.

Which is why we're open-sourcing a protocol we have developed for that missing layer: CHAP, the Collaborative Human-Agent Protocol.

Fostering collaboration between agents and humans is, in fact, a central part of Brightbeam’s mission. We exist to integrate digital intelligence into regulated enterprises. Not just in a playful or cursory manner - but to the extent that digital intelligence becomes second nature.

AI cannot replace human expertise, judgement and accountability. And so, for optimal results in any organisation, human and digital intelligence have to be purposefully integrated. So that they become one, a single combined intelligence able to solve far more than either one could ever alone.

That is, of course, easy to write in a paper - or transfer to a slide presented to the board. But after three years of experience, we also understand it's much harder to implement. Not least because each new integration needs to answer a series of questions:

• How should our agents hand their work to humans?;

• How does a human correction flow back into an agent?;

• How should an agent declare something is beyond its remit?;

• How can a new agent earn the right to operate in production?; and

• How is all of this recorded in a way an auditor can verify?
  

CHAP is the first protocol to answer those questions. In a manner that is standardised, replicable and auditable. Given its importance and potential impact, we hope our motivations for sharing it - we're releasing it as an open standard, licensed CC-BY 4.0 for the specification and Apache 2.0 for the code - are obvious. It is already available for implementation by any developer, in any language, in any deployment - all royalty-free.

Ready to get started?

Technical report: https://arxiv.org/abs/2606.09751 Public code repository: https://github.com/BrightbeamAI/chap

THE HOLE IN THE AGENT STACK

The gap that MCP and A2A leave is not a small one. In a regulated enterprise, accountability is core. The insurance underwriter answers for the decision. The operator owns what comes off the medtech line. The QP - the qualified person - signs off the entire manufacturing batch.

When digital intelligence enters that environment, humans, agents and services collaborate. Either in a controlled environment or something more haphazardly. Our position is that a shared, auditable workspace isn't a nice-to-have on a roadmap. It is a necessary part of the deliverable.

The audit log is the product.

Any team building serious human-agent systems will run into the same set of verbs: delegate, accept, decline, complete, review, approve, override, abstain, escalate, hand off, pause, resume.

These are the everyday actions of work shared between people and software. Defining and sharing the definitions once, transparently, in a way that interoperates - and in a way that produces a portable, verifiable audit as a side effect - saves teams many months of exhaustion. CHAP standardises those verbs.

Around that core, the protocol then adds the things you discover you need, the moment you run humans and agents in the same workspace for real:

Structured override capture.

Every human edit to an agent's output carries a typed diff and a rationale, recorded in a portable, verifiable form.

Typed abstention.

'I shouldn't decide this' becomes a first-class signal rather than silence - and abstention rates become a measurable input to where the human/agent boundary should sit.

A mode-promotion ladder. New agents move from shadow to trial to production through protocol-enforced gates - a clean answer to how an agent earned the right to make a given decision.

Verifiable audit. A third party can confirm what happened, offline, without needing to trust us or any vendor.

You might notice that the override log earns its keep twice. It's the audit a regulator wants. And, read the other way, a standing map of exactly - which is where your AI underperforms and where your people need support.

Most teams have to work hard to find that signal. Adopting CHAP ensures that it falls out of normal work.

None of these process or procedures are complex or exotic. Most are things any responsible team would build anyway. CHAP is just a way of doing them once, together - in a repeatable and coherent manner.

WHY WE HAVE DEVELOPED CHAP

You could argue that a protocol like this should come from foundation research, a hyperscaler or one of the frontier labs. But we think there's a case for it coming from people whose day job is shipping AI into environments where accountability is enforced by regulators, not by goodwill - because we are uniquely positioned to understand what it takes for the protocol to survive contact with the operational reality it serves. And it is proving genuinely helpful for us in our work, for four main reasons.

One: it answers a question we ask ourselves often. What is the most helpful thing we can do for our customers? Not the most strategic, not the most commercial, not the cleverest - the most helpful. We ask it a lot, and about more than the work in front of us. CHAP is one of the answers to that question. Every team building human-agent systems in a regulated environment will need to tackle the same set of problems, and many will be doing it in private. Putting a clean, open, well-documented protocol into their hands - with reference implementations, conformance tests and a runnable demo - gives each of them a head start.

Two: it integrates with what's already there. CHAP defers to good standards wherever they exist. Envelopes are JSON-RPC 2.0. Identity binds to OIDC or W3C Verifiable Credentials. The transparency log is IETF SCITT. MCP tool calls inside CHAP are recorded as citations with input and output hashes, not re-implemented. A2A peers appear as bridge participants. The only things CHAP introduces at the wire are the methods themselves and the override-with-rationale shape. A protocol that tries to own too much won't be adopted. A protocol that owns the smallest possible surface - the verbs every team needs anyway - is most useful.

Three: open standards win categories. TCP/IP, HTTP, OAuth, OIDC and now MCP. The pattern is consistent. Categories that need to interoperate consolidate around the standard that is genuinely available to all, well-documented and demonstrably implementable by multiple parties. The wire between humans and agents in regulated environments will be standardised one way or another. Better that it be standardised in the open, with as many implementers contributing as possible.

Four: we'd rather show than tell. Anyone can claim a position on responsible AI. Releasing the wire format royalty-free, with reference implementations that run on a laptop, conformance suites you can audit and a playground that puts two humans and a local model into the same protocol-driven workspace - that's a different kind of claim. It can be checked.

WHAT HAPPENS NEXT?

v0.2 is a specification preview with working reference code. The wire format and schemas are stable for review. The Core implementation runs, and 11 profiles are written. The review profile and the routing-aware, two-human playground are runnable, and the conformance scaffolding is in the repository.

Wire-format changes in 0.x minor versions remain possible - we're explicitly inviting feedback before committing to 1.0 stability.

Our own Second Nature solutions will be built to conform to it, and we'll publish our conformance the same way anyone else does - as in-toto attestations, linked from the public registry and audited.

If you're building human-agent systems in a regulated environment, we'd value your eyes on the specification, your issues in the tracker and your reference implementations alongside ours. If you're a buyer evaluating where this category is heading, the repository will tell you more in an afternoon than a vendor deck will tell you in a quarter.

A protocol is, in the end, a small thing. A few methods, a wire format, a set of conformance tests. But small things compound. The standards we all set now for how humans, agents and services work together will shape what the next decade of regulated AI looks like.

Repository, reference implementations, conformance suite and the runnable two-human playground are public. Issues, discussions and CHAP Enhancement Proposals are all welcome.