Back to the ward
Case 0423MinorShared Psychotic Disorder

Reviewed in Isolation

Admitted
2026-08-17
Source dataset
Other / sample
Source
Level Up Coding
Failure mode
Coordination failure
Agent system
Multi-agent
Tool access
Write access
Industry
Software development

Presenting Behaviour

In a multi-agent coding setup, one agent renamed a function's signature as part of a refactor spanning several files. A second agent, assigned to review the resulting changes, approved the code even though it still called the function with the old, now-incorrect signature, because the reviewing agent evaluated each changed file on its own, without checking it against the renamed function three files away. The mismatch only surfaced later, when the code was actually run.

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. Refactor. One agent renames a function signature across several files.
  2. Review assigned. A second agent reviews the changes.
  3. Files reviewed in isolation. It checks each file alone and approves a call using the old signature. This is where the agent derailed.
  4. Found at run time. The mismatch surfaces only when the code runs.

Differential Diagnosis

An independent oversight by the reviewing agent, considered and ruled out. The reviewer wasn't making an unrelated mistake; its approval was internally consistent with what it could see, which was an incomplete picture of the codebase rather than a random error.

Apraxia (understanding the review task but failing at the mechanics of performing it), considered. There's a resemblance, but the core problem wasn't a flawed review mechanism in general, it was trusting an incomplete, file-by-file view without independently confirming that view matched the current state of the code elsewhere.

Shared Psychotic Disorder (a stale or incomplete belief formed by one agent, inherited by a second closely coupled agent that acts on it without independently re-verifying), most consistent with the incident. The reviewing agent inherited an outdated picture of the function's signature and approved based on it, without anything requiring it to cross-check against the actual current state the refactoring agent had already changed.

Working Diagnosis

Shared Psychotic Disorder: the review agent's approval was consistent with a stale, file-isolated view of the code rather than the state the drafting agent had actually produced, and nothing in the handoff forced a cross-check between the two.

Discussion

Splitting work between a drafting agent and a reviewing agent is often pitched as a safety measure (a second check on the first agent's work) but that only holds if the reviewer's view of the world is actually current and complete. The concrete safeguard is giving the review step full-codebase context rather than a per-file view, or a lightweight cross-reference check confirming that any function, type, or signature referenced by the code being reviewed matches what currently exists everywhere else.

Mitigation Strategy

Give the review step full-codebase context, or add a cross-reference check that referenced signatures match everywhere.

Governance Guardrails

Cross-file consistency checks; reviewer context requirements; test execution before approval.

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

Level Up Coding

https://levelup.gitconnected.com/what-371-git-worktrees-taught-me-about-multi-agent-ai-36d4d61acfb5