When the agent gets it wrong
The report looked useful, but it had stopped reading an important source.
Earlier versions of my morning-report workflow carried an explanation that notes written by my note-taking app could not be read directly because of the machine’s sandbox. The workflow treated that explanation as settled and substituted earlier morning reports for the current weekly note.
That substitution missed live content. The failure was recorded in the workflow’s own history, and a later check found that direct file reads worked. The earlier diagnosis was wrong, or the restriction had changed. Either way, carrying it forward without checking made the report less accurate.
- AssumeAn old access diagnosis becomes a standing rule
- SubstituteEarlier reports replace the current source
- MissCurrent work never enters the new report
- CorrectAttempt the read and account for the evidence
See the consequence
The failure above is documented. The task names below are fictional examples that illustrate how it can happen.
Yesterday’s report
Open task: Arrange the workshop tour.
This tells us what yesterday’s report believed was open. It does not establish what is open now.
Today’s weekly note
Tour is arranged. New task: Prepare the equipment checklist.
Reading this source changes both parts of the work list: one task is finished and a new task needs attention.
A report built from the substitute
Open task: Arrange the workshop tour.
The output can be neatly formatted and still omit the new work. Repeating an old report gives it no new evidence.
The corrected rule
Attempt the source read before declaring it unavailable. If it fails, record the actual failure and make the gap visible. Check current evidence before surfacing an old task as open work.
The real workflow also requires its declared integrations to pass a preflight. A missing required connection stops report generation. Other source reads must have contribution evidence or an explicit accounting of why they did not contribute.
What changed
The workflow now explicitly tells the agent to try the read, use the current notes, and avoid reconstructing the task list from old reports. It also requires source accounting so that a missing input cannot disappear behind a polished summary.
Those are workflow rules, and an agent can still fail to follow them. I would not call the problem impossible now. What improved is that the expected behavior and the evidence needed to check it are written down.