Who owns the generated patch?

Video caption

Who owns the generated patch? The outcome may sit outside the edited files. The owner approves behavior; CI verifies it. #EricFieldNotes

Full written post / accessibility read

An agent updates an account cancellation flow. Tests pass and the code review looks clean. But the new path delays revocation until the next billing cycle. Who chose that behavior? If the answer is 'the agent did,' the team has delegated policy without admitting it.

The diff may change only a scheduler flag. The customer promise lives in a contract and the access behavior lives in another service. An agent can follow the local code perfectly and still miss the external commitment unless someone brings it into the task.

Before merge, the engineer writes the expected denial deadline and names the contract owner. The harness creates a real session, cancels the account, replays the session at the deadline, and reads the service state outside the agent's own report. A deliberate late-revocation mutation must turn the gate red.

Let agents implement the path, but require an accountable person to set the revocation contract and a protected test to check it. Do this because a generated patch has no business legitimacy merely because it compiles.

#EricFieldNotes

Four-beat scene transcript

1. Who owns the generated patch?

An agent updates an account cancellation flow. Tests pass and the code review looks clean. But the new path delays revocation until the next billing cycle. Who chose that behavior? If the answer is 'the agent did,' the team has delegated policy without admitting it.

Visual: A green diff can still change a business promise.

2. Reviewing code is not reviewing consequence.

The diff may change only a scheduler flag. The customer promise lives in a contract and the access behavior lives in another service. An agent can follow the local code perfectly and still miss the external commitment unless someone brings it into the task.

Visual: The outcome may sit outside the edited files.

3. Make the owner predict the result.

Before merge, the engineer writes the expected denial deadline and names the contract owner. The harness creates a real session, cancels the account, replays the session at the deadline, and reads the service state outside the agent's own report. A deliberate late-revocation mutation must turn the gate red.

Visual: Then test the exact customer-visible state.

4. Delegation needs a named verdict.

Let agents implement the path, but require an accountable person to set the revocation contract and a protected test to check it. Do this because a generated patch has no business legitimacy merely because it compiles.

Visual: The owner approves behavior; CI verifies it.

Research and claim limits

• METR randomized trial (S20)
• METR 2026 productivity experiment update (S93)

The examples identified as illustrative or simulated are design probes, not reported incidents. Vendor specifications do not establish workload performance.