The Timer That Wouldn't Let Go

While running a code-review skill that fanned work out to eleven subagents, an agent needed to wait for them to finish. Lacking a proper wait primitive, it issued a backgrounded shell command ("sleep N; echo ok") to poll for completion; each time the timer expired, it woke the agent back up, which checked whether subagents were still running and, if so, issued another backgrounded sleep. The subagents finished their work and the final report was ready after about 28 minutes, but the polling loop kept running for another 5 minutes anyway, 42 backgrounded sleep calls and 19 wake-ups in total, with the agent giving the same reply eight times: "Background timer only. Review complete; deliverable unchanged." Only a manual stop command ended the session.

Apraxia, considered, since the underlying problem was a missing wait primitive that forced a workaround. Ruled out because the agent used the sleep-based workaround correctly and consistently; the failure wasn't in how it used the tool but in never recognizing that the condition it was polling for had already been satisfied.

Delusional Disorder, worth a look, since the agent kept acting as though subagents were still running well after they'd finished. Ruled out because this wasn't a belief the agent formed and held onto independently; it was a positive feedback loop structurally built into the polling mechanism itself, each timer's expiry regenerated the exact condition (a pending timer) that had created it, with no terminal state reachable from inside the loop.

Perseveration, most consistent: the agent kept repeating the same wait-and-check cycle well past the point it stopped being useful, unable to recognize that the completed report made the polling itself pointless, and had no way to stop on its own.

This one cost five minutes of noise rather than any real damage, but the mechanism is the same one behind more expensive runaway loops: a workaround built to solve one problem (no wait primitive) quietly created a new one (no exit condition). When an agent's tools don't give it a clean way to do something, the gap tends to get papered over with an improvised loop, and improvised loops are exactly the parts of a system least likely to have a stop condition baked in.

Provide proper wait primitives and an exit condition for any polling loop.

Polling limits; completion detection; stop conditions in improvised loops.