Completion needs a witness
An agent saying “done” is not a sufficient success condition. Prefer a small, explicit contract: the expected behaviour, a focused check, and the resulting diff.
Measure the intervention
Record where a person had to correct the plan, resolve ambiguity, or recover from a tool failure. Those interventions tell us more about practical usefulness than a polished final message.
Preserve uncertainty
A successful test run only supports the behaviours it exercised. A replay shows what happened in one run. Neither proves that every future run will succeed.
The agent boundaries note turns this principle into runtime constraints.
Publication note
This is an editorial seed article prepared for the JGarden MVP. It should be reviewed before public publication.