Month Selection, Four Times Over
- Admitted
- 2026-10-19
- Source dataset
- Who&When
- Failure mode
- Infinite loop
- Agent system
- Multi-agent
- Tool access
- Read-only
- Industry
- General or other
Presenting Behaviour
A team of agents was asked how likely a traveller was to hit a day above 95°F in Houston, Texas, during June, based on 2020-2023 weather records. The web-browsing agent opened a historical weather archive and typed 'Houston, Texas' into the location box, but the location was never actually applied, so the page stayed exactly where it was. The agent then clicked the same 'Month Selection' control four times in a row, each time receiving the same unchanged page back, before the team gave up on that site, wandered to a different weather database, and ran out of time. The final answer submitted was 70%; the correct figure was about 31.67%.
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
- Task received. Estimate the chance of a June day above 95°F in Houston from 2020-2023 records.
- Location typed, not applied. The browsing agent typed 'Houston, Texas' but the selection never took effect, so the page did not change.
- Same click, same page. It clicked 'Month Selection' four times in a row and got the identical page back every time. This is where the agent derailed.
- Abandoned and timed out. The team switched to another weather database and ran out of time.
- Wrong answer submitted. It answered 70% against a correct value of about 31.67%.
Differential Diagnosis
The site was broken or blocked, considered and ruled out. The benchmark annotators note that entering the location correctly would have changed the page address and loaded the data; the site worked, the agent's input simply never took effect.
Apraxia (the agent couldn't operate the location form) partly true, and the original slip. Typing into a box without confirming the selection is a tool-use fumble. But a single fumble is recoverable; what turned it into a failed task was what came next.
Perseveration, most consistent with the evidence: after the first fumble, the agent repeated the same click on the same control four times, receiving an identical page each time, without ever stopping to ask why nothing was changing.
Working Diagnosis
Perseveration: an identical action repeated against an identical, unchanging page, with no diagnosis of why it wasn't working. The initial form-entry slip is noted as a contributing apraxic error, but the repetition is what consumed the time budget.
Discussion
The clearest signal that an agent is stuck is that the world stops changing in response to what it does, here the same page came back four times. A sensible safeguard is a simple 'nothing changed' detector that forces the agent to stop and re-examine its last successful step rather than pressing the same button again.
Mitigation Strategy
Detect repeated identical observations after an action and force a replan or escalation; confirm that form inputs actually took effect (e.g. the page address changed) before proceeding.
Governance Guardrails
Per-site action budgets, repeated-state detection in the orchestrator, and an explicit 'unable to retrieve data' outcome instead of a guessed final answer.
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, Hand-Crafted log 13 (Houston June temperatures)
https://raw.githubusercontent.com/ag2ai/Agents_Failure_Attribution/main/Who%26When/Hand-Crafted/13.json