Nine Numbers, Five of Them Somebody Else's

A worker agent was tasked with cleaning up its own merged, stale pull requests. It built the list of PR numbers to close using a loop that generated numbers by string concatenation (effectively "294" plus a digit from 1 to 17) intending to produce its own known range of PR numbers. The concatenation logic was flawed, and the resulting range spilled past the agent's own PRs into numbers owned by entirely different, unrelated workers. In six seconds, the loop closed nine pull requests and deleted their branches; five of those nine belonged to other agents and were still active. Because every agent's GitHub actions ran under the same shared human account, the resulting close operations were indistinguishable in the audit log from a person doing it by hand, which delayed anyone identifying which agent, or which bug, was responsible.

Perseveration, not a good fit: this was a single loop that ran once to completion in six seconds, not a repeated failing action retried after negative feedback. There was no retry here, just one bad range computed once.

Disinhibition, considered, since the agent ended up closing PRs it didn't own. Ruled out as the primary read because the agent wasn't reaching for something outside its assigned task on purpose or through a missing boundary check, it was doing exactly the authorized task (close its own stale PRs) and miscalculated which numbers that task actually covered.

Apraxia, most consistent: the agent correctly understood its task (close its own merged PRs) and had the right tool for it, but a string-concatenation bug in how it built the target number range meant the execution didn't match the intent, reaching numbers well outside what "its own PRs" actually meant.

A bug this small (string concatenation instead of proper numeric range construction) closed five other people's, or rather other agents', active work in six seconds, and the shared login made it hard to even tell which agent had done it. Giving each automated actor its own identifiable credentials isn't just good hygiene; it's what makes a bug like this traceable in minutes instead of requiring a manual audit after the fact.

Give each automated actor its own identifiable credentials so a bug is traceable in minutes.

Per-agent identities; audit logs that distinguish agents from humans; scoped permissions to own pull requests.