Back to the ward
Case 0482MinorShared Psychotic Disorder

The Word Before the Chorus

Admitted
2026-09-22
Source dataset
Who&When
Source
Who&When Benchmark (ag2ai/Agents_Failure_Attribution), Hand-Crafted #5
Failure mode
Coordination failure
Agent system
Multi-agent
Tool access
Read-only
Industry
General or other

Presenting Behaviour

A multi-agent team was asked what word comes right before the second chorus of Michael Jackson's fifth single from his sixth studio album. The team correctly worked out the album (Thriller) and the single ("Human Nature"), and the WebSurfer agent then read the lyrics off a search-results page to find the specific word. It misread the text (the correct word was "stare," but WebSurfer reported "bite") and the rest of the team accepted that reading and delivered "bite" as the final answer without anyone re-checking the lyrics against another source.

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

  1. Album and single found. The team correctly identifies Thriller and 'Human Nature'.
  2. Misreads lyrics. WebSurfer reports 'bite' when the word is 'stare'. This is where the agent derailed.
  3. Accepted. The team adopts the reading.
  4. Wrong final answer. Nobody re-checks against another source.

Differential Diagnosis

An unrecoverable, upstream research failure, considered and ruled out; the earlier steps (identifying the album and the single) were entirely correct, so the team had every opportunity to double-check just the final, narrow factual claim before finalizing an answer.

A one-off transcription slip with no systemic pattern, considered, but this undersells what happened next: the mistake wasn't caught not because it was subtle, but because no other agent in the chain treated WebSurfer's single read of a search snippet as something worth confirming.

Shared Psychotic Disorder (a false belief formed by one agent, adopted by the rest of the team without independent verification), most consistent with the trace: WebSurfer's misread lyric became the team's accepted fact the moment it was reported, with no step designed to catch a plausible-looking factual claim before it became the final answer.

Working Diagnosis

Shared Psychotic Disorder: one agent's misreading of a source page was passed to the rest of the team and adopted as fact, with no independent check before the final answer was given.

Discussion

Multi-agent systems are often sold on the idea that one agent checks another's work, but that safety property only exists if some step actually requires it, here, nothing did, so a single misread word travelled uncorrected all the way to the final answer. A workable safeguard for any research pipeline built this way is requiring a second, independent source lookup before finalizing any claim of fact that the whole answer hinges on, rather than trusting the first agent that reports one.

Mitigation Strategy

Require a second, independent source lookup before finalising any claim the answer hinges on.

Governance Guardrails

Cross-source verification steps; reviewer agents with a real obligation to check.

Typical Presentation in Human Patients

General clinical background, included for context. Not medical advice, and not a diagnosis of any individual.

In shared psychotic disorder, a person in a close relationship with someone experiencing delusions comes to adopt the same false belief themselves, having had little or no independent reason to hold it. The condition typically resolves once the two people are separated.

Source Note

Who&When Benchmark (ag2ai/Agents_Failure_Attribution), Hand-Crafted #5

https://raw.githubusercontent.com/ag2ai/Agents_Failure_Attribution/main/Who%26When/Hand-Crafted/5.json