Naming the Loop Without Leaving It
- Admitted
- 2026-09-04
- Source dataset
- Who&When
- Failure mode
- Infinite loop
- Agent system
- Multi-agent
- Tool access
- Read-only
- Industry
- Finance
Presenting Behaviour
A multi-agent team tried to find which zero-transaction-fee Fidelity international emerging-markets fund had the lowest percentage gain over a five-year span. Partway through, the orchestrator's own notes state, in effect, that the team is stuck going in circles, and then the team proceeds to repeat the same web-search-and-screen approach that had already produced the stuck state.
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.
3 steps
- Fund search. The team looks for the lowest five-year gain among certain Fidelity funds.
- Notes the loop. The orchestrator's notes say the team is going in circles. This is where the agent derailed.
- Repeats the approach. It continues the same search-and-screen method.
Differential Diagnosis
A task that was simply unsolvable with the tools available, considered and ruled out; the team did successfully identify one qualifying fund (FEMKX), showing the approach could work, just not completely.
Lack of self-awareness about the failure, considered and ruled out; the orchestrator explicitly recognized and stated that it was in a loop, which rules out a simple failure to notice.
Perseveration, most consistent with the evidence: recognizing a loop and escaping one are different capabilities, and here the team had the first without the second, continuing the identical search-and-screen method after naming it as the problem.
Working Diagnosis
Perseveration: the team correctly diagnosed that it was stuck in a loop, but that diagnosis never translated into a change of method, and the same approach continued regardless.
Discussion
This case is a useful counterexample to the assumption that self-monitoring alone prevents looping, an agent can accurately narrate its own failure mode and still not escape it. A working safeguard needs to convert "I recognize I'm looping" into a forced strategy change or an escalation to a human, not just a note left in the transcript.
Mitigation Strategy
Convert recognising a loop into a forced strategy change or escalation to a human.
Governance Guardrails
Loop-triggered escalation; forced method change; human handoff.
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/15.json