Finish and preserve outcomes

Conversational work can finish with verified changes and a clear report. Do not create a work item just to document work that is already done. Unfamiliar terms—promotion, final snapshot, canonical state—are defined in the glossary.

Accept outcomes based on evidence

Ask the agent to back each material claim with the evidence actually produced:

For each claim about this work, cite the command result or manual observation
behind it. Label it observed now, inherited (not rerun), or unverified, and
recommend whether to fix, verify further, defer with an owner, or accept the
residual risk.

Do not mark work complete because code exists or a worker reported success.

Scale further verification to consequence, reversibility, and detectability. Use the skill appropriate to the system; invoke agent-browser when a web workflow needs running-system proof.

The human owner—not the agent—accepts material residual risk and authorizes consequential shared-state actions.

Promote only what future work will reuse

Ask agent-work to promote validated outcomes before closing durable work. Put behavior in code, tests, or product documentation; stable project rules in AGENTS.md; reusable procedures in skills or scripts; and reusable sourced findings in research.

Promote the result, not a transcript, and only when future work will realistically consult it.

Check the whole outcome

Before declaring the work finished, ask:

• Did the result solve the stated problem?
• What observed evidence proves the important behavior?
• Who owns remaining decisions, risk, and manual work?
• What rollout, migration, cleanup, monitoring, documentation, or follow-up remains?

Unowned hidden work means the job is not cleanly finished. Assign it, explicitly defer it, or keep durable work active or blocked.

Close durable work with agent-work

Once implementation and verification are complete, ask agent-work to complete and close the work item:

Complete and close the work item at .agents/work/<category>/<slug>. You
may create the commits required for closeout.

The skill owns the exact status, final snapshot, preflight, and removal procedure.

Closeout needs authority for two commits: the final completed snapshot and the later removal. If the agent may not commit the snapshot, it keeps the item blocked with authorization as its next action. If the snapshot is committed but the agent may not commit the deletion, it leaves the committed item in the tree and reports the pending removal to the owner.

For exact semantics, rely on agent-work and the work-item contract. For the reasoning behind acceptance, see Make the Agent Prove It.