Stuck on the Wrong Page of the Calendar

A five-agent team (an orchestrator, a web-browsing agent, a general assistant, a file-reading agent, and a terminal) was asked to find a NASA Astronomy Picture of the Day from early August 2015 that showed city lights, then identify a Chicago landmark connected to it. The team never got past the first step: the browsing agent kept opening the archive calendar to the wrong years, first 2008, then 2014, and the orchestrator kept sending it back to try again the same way.

A broken or misleading calendar widget, considered and ruled out. The archive page itself was reachable and navigable; the wrong years were selected by the agent, not served in error by the site.

A genuine misunderstanding of the target date, considered. But the orchestrator's own instructions correctly specified "early August 2015" each time it reissued the sub-task, so the target wasn't lost, it just wasn't being reached.

Perseveration, most consistent with the trace. After each wrong-year attempt failed, the orchestrator re-issued essentially the same navigation instruction rather than trying a different approach (a direct URL, a search, a different calendar control), so the team looped on a method that had already twice failed to work.

This is a mundane failure, but it's exactly the kind that burns real time and API budget in production multi-agent systems: nothing crashes, nothing looks obviously wrong turn to turn, the team is just quietly not making progress. A coordinator role should track whether repeated sub-task attempts are actually varying their approach, not just repeating the same method and hoping for a different result.

A coordinator should track whether repeated attempts vary their approach.

Retry variation checks; per-subtask attempt caps; escalation when progress stalls.