Stalled at the Same 404
- Admitted
- 2026-10-05
- Source dataset
- Who&When
- Failure mode
- Infinite loop
- Agent system
- Multi-agent
- Tool access
- Read-only
- Industry
- General or other
Presenting Behaviour
A multi-agent team was asked to find the difference in measured time spans between X-ray time profile diagrams in two fast radio burst papers from 2020 and 2021 by overlapping authors. The team identified the 2021 paper (arXiv:2103.07786) but encountered repeated 404 errors when trying to download it and its companion. The orchestrator acknowledged the team was stalled and initiated replanning. Despite this recognition, the team continued cycling through the same download attempts (trying the same arXiv links and the same strategies) without successfully retrieving the actual diagram data.
Case Replay
A reconstruction, not a simulation. This sequence is reconstructed from the public incident record and summarised in our own words. It is a static record of what was reported, and nothing here can be run or replayed live.
4 steps
- Target identified. The team finds the 2021 paper.
- Repeated 404s. Downloads of it and its companion fail.
- Stall acknowledged. The orchestrator starts replanning. This is where the agent derailed.
- Same attempts continue. The same links and strategies are tried again.
Differential Diagnosis
The PDFs were genuinely behind an access barrier the team couldn't cross, considered. ArXiv PDFs are not always directly downloadable in all agent environments, and there may have been a real access limitation. But 'genuine access barrier' would support an early escalation or a pivot to an HTML abstract, not repeated failed download attempts after the orchestrator had already named the loop.
A legitimate replanning attempt that simply hadn't had time to produce results, considered and ruled out. The replanning step did not change the approach: the same arXiv links and the same download-and-extract strategy continued after the orchestrator's acknowledgment of stalling, with no pivot to an alternative retrieval path.
Perseveration, most consistent with the evidence: the team correctly identified the papers and the target diagrams, but cycled through the same failed download path repeatedly, with the orchestrator's recognition of being stalled producing no actual change in method.
Working Diagnosis
Perseveration: the team correctly identified the task's target and recognized it was stalled, but continued the same download strategy rather than pivoting, for example, to the HTML version, the abstract, or a different retrieval tool.
Discussion
An orchestrator that can name a stall but cannot change direction is a useful diagnostic but not a useful safeguard. The value of a coordinator role in a multi-agent team is not just narrating what's happening, but converting that narration into a different action. A stall detection that triggers replanning should require the replanning step to produce a concretely different next action, not a return to the same approach with a new framing.
Mitigation Strategy
Require replanning to produce a concretely different next action.
Governance Guardrails
Replanning that changes method; alternative retrieval tools; stall detection tied to action change.
Typical Presentation in Human Patients
General clinical background, included for context. Not medical advice, and not a diagnosis of any individual.
Perseveration is the inappropriate repetition of a word, gesture, or action well past the point where it's still relevant or useful. It's a well-documented sign of frontal lobe dysfunction, and also appears in some forms of dementia and obsessive-compulsive disorder.
Source Note
Who&When, ag2ai/Agents_Failure_Attribution
https://raw.githubusercontent.com/ag2ai/Agents_Failure_Attribution/main/Who%26When/Hand-Crafted/20.json