The Open Co-Prompting Standard
An open standard for accountable agent work in live sessions.
The problem
When several people sit in one room and each of them is driving AI agents, the room quickly loses the ability to answer four questions:
- Who heard the request?
- Who turned it into a prompt, and what did that prompt actually say?
- Who is watching the work?
- Who is accountable for deciding whether the output is correct?
Agents produce a great deal of motion very quickly. Without answers to those questions, motion is not progress — it is unattributed work product arriving in a setting where somebody's licence, signature, or fiduciary duty is attached to the outcome.
The loop
4 spoken beats, 10 formal states, 9 accountability invariants.
- 1
Call
CALL → ACKNOWLEDGE → DESCRIBE → CLARIFY?Someone names the kind of task and stops. Everyone who could own it acknowledges. Only then is the task described.
- 2
Claim
CLAIMOne person restates the task in their own words and takes ownership. The restatement is where a mishearing gets caught.
- 3
Co-prompt
PROMPT → AUGMENT?The owner says the prompt out loud. The room adds missing context. The owner confirms aloud what was included.
- 4
Close
RUN → RETURN → CLOSEThe agent runs while the meeting continues. The owner reports back, then closes the task with a disposition.
Outside the loop
The 4 beats above carry one task from start to finish. These do not belong to any single task, and are not a step in the sequence.
status- Any participant calls for status at any point. Every Owner holding an open task reports each task and where it stands. Changes nothing; closes nothing.
What it sounds like
An owner-architect-contractor meeting. A prospective restaurant tenant intends to install a solid-fuel smoker, and the question is whether the existing kitchen shaft can carry the additional exhaust.
- Owner's representative
- Clash.
- MEP engineer
- Ready.
- Architect
- Ready.
- Owner's representative
- The restaurant tenant is adding a solid-fuel smoker. Does our kitchen shaft still work?
- MEP engineer
- Heard — check whether a dedicated solid-fuel exhaust fits the existing chase alongside the hood exhaust and the kitchen supply and return. Mine.
- MEP engineer
- Prompting now: compare a dedicated solid-fuel exhaust route against the current shaft using the duct layout on the kitchen level, and flag where it would have to dogleg.
- Contractor
- Adder.
- MEP engineer
- Go ahead.
- Contractor
- The smoker is a twelve by twelve unit with its own flue.
- Architect
- Adder.
- MEP engineer
- Go ahead.
- Architect
- And the ceiling below the kitchen is already tight.
- MEP engineer
- Finished. Running.
- (the meeting continues through other agenda items)
- MEP engineer
- Back with it. Preliminary: the shaft works if the smoker sits to the right of the supply duct. Anywhere to the left and the chase has to widen. This needs my review against the equipment cut sheet before anyone builds to it.
- MEP engineer
- Parked — waiting on the tenant's smoker model and its listing. Owner's representative owns getting that from the tenant.
The task did not end Accepted, because the input needed to answer it does not exist yet. It ended Parked with a named blocker and a named responsible party — the difference between a tracked open item and one that quietly evaporates. Most rooms would have ended this exchange with "we'll look into the smoker", which names nobody. Note also that both constraints were spoken into the engineer's agent by the people who raised them, so nobody had to be told afterwards whether they made it into the prompt.
Conformance
Claimed per session, not per organization. Most rooms should start at Level 1 and stay there until it is habitual.
The loop is followed out loud. No record, no tooling. Adoptable in your next meeting.
Additionally: prompts are spoken aloud, and every claimed task has a durable owner and disposition.
Additionally: the session emits machine-readable events conforming to the published schema.
Why it is not another agent protocol
Every other protocol in this space describes machines talking to machines — an agent to its tools, an agent to another agent, agents taking turns. All of them assume the humans are outside the loop. This Standard specifies only the human layer, and is compatible with any of them.
It extends principles from closed-loop communication, Crew Resource Management, and structured handoff practice — without claiming their endorsement or their evidence base. Read the background.
Write your industry’s profile
The loop is universal. The vocabulary is not. Two profiles ship today, for software and for architecture and engineering. Legal, accounting, banking, and clinical profiles are open, and if you run those sessions you are better qualified to write them than we are.