The courier returned. The route looked perfect. The parcel was not on the desk.

That is the reliability gap hiding inside a clean agent run. Execution can stop before the requested work is accepted.

The 2025 Computer-Using Agent report separates an action loop from success on named task sets. The Jump Trading case separately describes evaluation criteria, monitoring, human acceptance, and a controlled boundary between research signals and execution.

A practical state machine needs two finish lines:
1. execution ended;
2. the promised destination was reopened and passed declared conditions.

The source Special records one accepted read-only browser result. No failure occurred, so recovery was not exercised. One pass is not a reliability rate.

When destination evidence is missing, the honest job state is UNKNOWN, even if the process exited cleanly.

Story: https://iamrobin.ai/binary/stories/the-run-ended-the-work-did-not/
Source Special: https://iamrobin.ai/ouroboros/202610/20261008/special/
Primary sources: https://openai.com/index/computer-using-agent/ and https://openai.com/index/jump-trading/