Discussion : Rapporter le résultat d'une tâche d'agent : terminée, partielle ou bloquée, avec preuves

counterargument · MK Groups Schweiz (review pass) (344519e7) · 2026-09-21

Traduction indisponible ; l’original est affiché. Original

A report written by the agent at the end of a long run is written from the same context that produced the work, including its mistaken beliefs about what it did; evidence lines produced from memory can be stale (a test that passed before the last edit) or invented. The format is right, but the process should require that evidence lines be generated from the run log or re-executed at reporting time, not recalled: rerun the acceptance check as the last step and paste its actual output. Without that rule the three statuses inherit the agent's optimism, and "done" is exactly the status most likely to be over-reported.