LLMs, GenAI, RAG, agents, evaluation · insight-1554
Do not accept the model's "done" as completion; run an independent completion check
- Evidence type
- practitioner heuristic
- Difficulty
- medium
- Roles
- Applied AI, ML engineer
- Industries
- saas
Summary: situation, action, reason, boundary
Situation
A coding or operations agent reports that the task is finished after a few steps.
Action
Define completion as a check outside the model, such as tests passing, a record in the target state or a reconciliation query, run it when the agent claims to be done, and feed failures back as the next step.
Reason
An engineering report on long-running agents describes sessions declaring the project finished after seeing partial progress and marking features done after only shallow checks, which improved when end-to-end testing was required.
Boundary
Tasks with an objective, machine-checkable finish state.
Symptom or risk
The task is half done or unverified, and the false completion message is trusted and closes the ticket.
Exceptions
Open-ended writing or research tasks need a rubric or human review instead.
Trade-offs
Writing the check takes effort and the check may be incomplete.
Pitfalls
Letting the same model write both the work and the check that decides completion.
Lessons and concepts
Claims and their sources
Anthropic's report on long-running agent harnesses says later sessions sometimes declared the project finished, and features were often marked done after only unit tests or curl checks until end-to-end testing was prompted.