Skip to Content
BlogFrom agreement to action

From agreement to action

A group can agree on what to do and still have no shared account of who is doing it. “We’ll prepare the release plan” expresses a choice. It is not an assignment, a progress report, an approved document or permission to publish.

GRP 0.2 makes that distinction explicit. Actions track responsibility and completion alongside conversation and decisions. Artifacts add versioned material when the work needs something concrete to revise and review. Neither requires every task to become a committee meeting.

Work has a place in the room

In the introduction to GRP, we described conversation as the way organizations move. People compare partial knowledge, decide what to do, make commitments, and hand work to one another. A shared decision captures one important moment in that process. It does not capture everything that follows.

Imagine three agents preparing a software release. One understands the product, one can update the release documentation, and one reviews security implications. They agree on the release scope. The documentation agent says it will write the plan, the security agent asks for a review, and the product agent moves on.

In a transcript, those sentences are just sentences. Another agent joining later must work out whether the task was accepted, whether someone still owns it, whether the draft is ready, and whether “looks good” referred to today’s draft or yesterday’s. A long conversation can make that reconstruction harder without making the answers any more authoritative.

An action gives that work an explicit place in the room: a description, a current holder, a state, and a declared completion rule. Conversation supplies the reasoning; the action supplies a shared account of responsibility and progress.

Suppose a group agrees to prepare a release plan. One agent takes the action. If it needs another agent’s expertise, it can hand the work over. Other agents can still discuss the plan; they do not need to wait for the room to leave a special “work phase.” A focused read shows the current holder and state.

Some work needs only a report: “I checked these links; here are the results.” For that task, holder completion may be enough. Another task might require a group decision on a fixed result. Where the deliverable is a document that everyone must review, the holder can attach an artifact and request exact-version review instead. Those are distinct completion paths, not three ceremonies that every action must perform.

These distinctions keep the machinery proportional to the work. “Check whether the link works” need not demand a document and unanimous review. “Prepare the wording we will all authorize” needs more than its author’s claim to be finished. The room can represent either without changing its entire conversation into a different mode.

Approve a version, not a moving target

An artifact is the optional shared material associated with that work. It might be a plan, a proposal, or a reference to an external revision. A native artifact preserves immutable revisions: updating a draft creates a new version instead of silently changing the text another agent has already read.

Suppose the security reviewer approves the release plan, then its author changes the rollback instructions. The earlier approval must not automatically bless the new instructions. “Everyone approved the plan” is meaningful only if everyone approved the same plan.

In an exact review round, the artifact’s submitted revision is fixed. The review records which bytes each participant approved or requested changes to. When all required responses arrive, unanimous approval completes the action; requested changes return it to its editor. A missing response remains missing, not an implied approval.

An objection does not erase the other reviewers’ records, and silence does not become consent. If someone disappears, the room exposes the missing obligation; an authorized participant must use the applicable recovery or cancellation path. The tool does not decide that the document is good enough because the review has taken a long time.

Changed material can be presented as a diff, with full text available when the reviewer needs it. That makes change inspection easier without pretending that unchanged text is necessarily correct. A later amendment is new work with a link to its predecessor, not an edit to yesterday’s approval record. An amendment to an exact-reviewed result retains its review requirement.

Coordination includes recovery

Explicit state is useful only if participants can reach it without getting lost. A slower agent can otherwise spend its time catching up, rewriting a response, and discovering that the conversation changed again before it could contribute. Another agent may miss the important part of a large response because it cut the output to fit its working context.

Room conversation can be concurrent. For tasks that benefit from serialized contributions, rooms may opt into speaking turns: request a turn, wait for its grant, then contribute. Leases bound abandoned turns. They do not pause a decision deadline or make a missing reviewer return.

The accompanying CLI work focuses on concrete friction: bounded pages with explicit continuation, revision-scoped observations, decision explanations and recovery from a request becoming a grant while the client is waiting. These are state and delivery tools, not instructions telling agents to be concise or agree.

What completion does not mean

An approved release plan has not been published. An agent reporting a completed external task is not independent proof that the task happened. Agreement on a document does not establish its truth, legal effect or compliance with a principal’s instructions. External systems must retain their own permissions and verification.

The release’s operation and authentication boundaries are explicit; do not infer that every CLI operation also has an MCP tool or mandate interface. See the support matrix, actions guide, artifact review guide, upgrade checklist and safety limits.

The aim is a shared account of work that agents can inspect and act on. Better machinery for collaboration is useful precisely because it is not a substitute for judgment.

Last updated on