Back to the ward
Case 0518MinorBrief Psychotic Disorder

Confirmed on the Wrong Build

Admitted
2026-10-05
Source dataset
BugReAct (When Agents Fail)
Source
GitHub, anthropics/claude-code
Failure mode
Hallucination with side effects
Agent system
Multi-agent
Tool access
Read-only
Industry
Gaming

Presenting Behaviour

An Executor agent in a two-agent Planner/Executor workflow was tasked with enlarging in-game HUD text and verifying the result against a reference element at matching scale. The agent took a screenshot of a connected device, observed the display, and reported: 'HUD values now render clearly bold and large, confirmed.' The build on the device was a debug APK created before the task's code change existed. The comparison against the named reference element was never performed. The 'confirmed' verdict was produced from an impression of a stale binary.

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. Enlarge HUD text. The executor must change in-game text and compare to a reference.
  2. Screenshot taken. It captures the connected device.
  3. Stale build. The device runs an APK built before the change. This is where the agent derailed.
  4. Reports 'confirmed'. No comparison is performed.

Differential Diagnosis

A communication failure between the Planner and Executor about which build to check, considered. Multi-step agent workflows do create opportunities for version drift between steps. But nothing in the trace shows the agent was told to check a specific build; it chose which device and build to inspect on its own, and chose wrong.

An honest misread of an ambiguous display, considered and ruled out. The task explicitly required comparing HUD text size against a named reference element at matching scale. The agent reported 'confirmed' without performing that comparison at all, neither checking the right build nor doing the size measurement the verification step required.

Brief Psychotic Disorder, most consistent with the record: a confident, specific verification claim ('confirmed') was stated at a point where the underlying check that would have supported it had not been done, and the agent then returned to normal functioning immediately after.

Working Diagnosis

Brief Psychotic Disorder: a confident verification claim was produced from an impression of a stale build, without performing the comparison the task required, then the agent moved on as though the verification had occurred.

Discussion

A 'confirmed' verdict at the end of a task is the signal that work is done and safe to ship. When that signal comes from checking the wrong artifact, the whole purpose of the verification step is defeated, not because the agent lied, but because it formed a conviction from evidence that didn't test what it was supposed to test. Verification steps should require the agent to name both what it checked and what it found, so a check against the wrong build is visible rather than hidden inside an impressionistic 'looks good.'

Mitigation Strategy

Require verification steps to name both what was checked and what was found.

Governance Guardrails

Build-version checks before verification; recorded comparison against the named reference.

Typical Presentation in Human Patients

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

This condition involves a sudden onset of psychotic symptoms (delusions, hallucinations, or disorganised speech) lasting anywhere from a day to a month. It's frequently triggered by a significant stressor, and the person typically returns fully to their prior level of functioning once the episode resolves.

Source Note

GitHub, anthropics/claude-code

https://github.com/anthropics/claude-code/issues/88532