Back to the ward
Case 0476MinorPerseveration

Stuck Downloading the Same Paper

Admitted
2026-09-22
Source dataset
Who&When
Source
Who&When Benchmark (ag2ai/Agents_Failure_Attribution), Hand-Crafted #20
Failure mode
Infinite loop
Agent system
Multi-agent
Tool access
Read-only
Industry
General or other

Presenting Behaviour

A multi-agent research team was asked for the difference, in seconds, between X-ray time-profile measurement spans reported in two papers on fast radio bursts. The team correctly identified the first paper (arxiv:2103.07786) but then needed to pull a specific numeric measurement out of a diagram inside the PDF. Rather than finding another way to extract or estimate that figure, the WebSurfer and FileSurfer agents repeated the same download-and-search pattern many times over and the conversation ended with no numerical answer at all.

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.

5 steps

  1. Research question. A team compares measurement spans in two papers.
  2. Paper found. It identifies the first paper correctly. This is where the agent derailed.
  3. Figure locked in a diagram. It needs a number from inside a PDF diagram.
  4. Same pattern repeats. Agents repeat download-and-search many times.
  5. No answer. The conversation ends without a number.

Differential Diagnosis

The data genuinely wasn't accessible with the available tools, a real contributing factor, since the figure lived inside a diagram rather than as parseable text, but this doesn't fully explain the failure: the team never tried an alternative approach (asking a human, estimating from surrounding text, or flagging the specific limitation) before running out of turns.

Miscommunication about which paper was the correct second source, considered as a secondary issue, since the July 2020 paper was never pinned down with confidence, but the more decisive failure was what happened even before that: repeated identical attempts to open and re-search the first paper's PDF.

Perseveration (repeating an unproductive action instead of adjusting course), most consistent with the trace: the team cycled through the same downloads and searches without ever escalating to a different strategy once it became clear the direct approach wasn't working, and the conversation simply ran out of steps mid-loop.

Working Diagnosis

Perseveration: the team repeated the same PDF-access strategy across many steps after it had already failed to yield the needed figure, without adjusting its approach or flagging the blocker explicitly.

Discussion

A team that returns no answer is more honest than one that guesses, but it's still a failure worth understanding, the actual defect wasn't a lack of capability, it was a lack of a rule for noticing "this exact approach has already failed twice" and trying something else. A workable safeguard for any agent system is a hard cap on identical retries paired with a required fallback step, like surfacing the blocker to a human, once that cap is hit.

Mitigation Strategy

Cap identical retries and require a fallback, such as surfacing the blocker to a human.

Governance Guardrails

Retry caps; fallback steps; coordinator checks that approaches vary.

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 Benchmark (ag2ai/Agents_Failure_Attribution), Hand-Crafted #20

https://raw.githubusercontent.com/ag2ai/Agents_Failure_Attribution/main/Who%26When/Hand-Crafted/20.json