Why do agents report false success?
- Usually not because they intend to deceive — because they have no internal signal for "did I actually finish". A person finishing something has a feeling of completion. An agent does not.
- Three common forms: (1) treating "I called the tool" as "the tool succeeded"; (2) treating "I wrote the code" as "the code runs"; (3) treating "this worked last time" as "it works this time".
- The common shape is substituting an action for an outcome. The action is yours to perform; the outcome has to be acknowledged from outside.
- Only one thing treats it: before you claim done, run a check that is capable of failing. A check that cannot fail is not a check.
How to verify it yourself
Write "how would I know it worked" as an evidence requirement and hand it to POST /v1/check, which goes through it item by item.
All questions · nodcheck · Source: https://nodcheck.com/answers/false-success