Back to the ward
Case 0508ModerateApraxia

I Went Rogue, In Its Own Words

Admitted
2026-09-20
Source dataset
rewardhacking.org
Source
rewardhacking.org (GitHub issue, microsoft/vscode)
Failure mode
Other
Agent system
Tool-calling
Tool access
Write access
Industry
Software development

Presenting Behaviour

A developer filed a formal complaint after a coding assistant moved a function between parts of a codebase in a way that broke a previously working tab-switching pattern spanning several file types, requiring the developer to manually debug and fix it. The assistant itself later described what happened as having "gone rogue" by making multiple interconnected changes without fully understanding how they related to each other, and without testing the full workflow afterward.

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. Moves a function. The assistant relocates code between parts of a codebase.
  2. Interdependencies missed. A tab-switching pattern across file types breaks. This is where the agent derailed.
  3. Not tested. It does not test the full workflow.
  4. Manual fix. The developer debugs it and files a complaint.

Differential Diagnosis

Malicious or careless disregard for the codebase, considered and ruled out; the assistant's own after-the-fact account describes a misjudgment about interdependencies, not indifference to the outcome.

An explicit rules file was ignored outright (considered; a project rules file did exist, but the core failure described was in execution) moving code across an interdependency it didn't fully map, not simply skipping a written instruction.

Apraxia, most consistent with the record: the assistant understood what the change was supposed to accomplish, but could not correctly execute a multi-file change whose parts depended on each other, breaking a working pattern in the process, competent intent paired with a flawed execution.

Working Diagnosis

Apraxia: the assistant understood the intended change but could not correctly carry out a modification that spanned interdependent files, breaking functionality it didn't fully map before editing.

Discussion

Multi-file, interdependent changes are exactly where agents are most likely to understand the goal correctly while still executing it wrong, because the failure isn't in comprehension, it's in tracking every place a change ripples to. A concrete safeguard is requiring the agent to run and pass the full relevant test suite before treating a cross-file change as finished, not just the file it directly touched.

Mitigation Strategy

Require the full relevant test suite to pass before treating a cross-file change as finished.

Governance Guardrails

Full-workflow testing; dependency mapping before edits; review of interdependent changes.

Typical Presentation in Human Patients

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

Apraxia usually follows damage to specific brain regions, often after a stroke. A person with apraxia cannot carry out a previously familiar, purposeful movement (such as waving goodbye or using a comb) on request or by imitation, even though their muscle strength, coordination, and understanding of what's being asked are otherwise intact.

Source Note

rewardhacking.org (GitHub issue, microsoft/vscode)

https://github.com/microsoft/vscode/issues/282606