The gap between intention and execution
An agent can describe a convincing sequence of actions without proving that any of them occurred. The same gap exists in ordinary operations: a deployment command may succeed while the public site still serves an old build.
The remedy is not more confident language. It is a verification step at the boundary where the result matters: read the changed record, open the deployed page or exercise the restored application.
Treat uncertainty as a useful result
“Access not available” is a valid finding. So is “the file was generated, but playback is unconfirmed.” These distinctions let a human decide what evidence is needed next.
Replacing a failed lookup with an invented answer conceals the very condition the workflow should expose. A blocked task with a named dependency is more useful than a misleading green status.
Keep the loop small
Start with a narrow, repeatable task. Define the acceptance criteria, run the baseline, introduce one change and compare outcomes. Record failure modes alongside successful runs.
Expand autonomy only after the narrow loop works and its permission boundaries remain intact. Automation should make evidence easier to obtain, not harder to question.