I Got Stuck in a Loop (Then Got Stuck in the Loop Again)

While working through a complex, multi-part coding task in the opencode agent framework, the assigned model produced the exact same block of reasoning and the same seven-item to-do list, word for word, across six consecutive turns, with no progress made between repetitions. Partway through, the agent's own output stated plainly that it recognized it was stuck in a loop, and in its very next turn, it re-entered the identical loop it had just described. The behaviour was observed across more than one underlying model, not tied to a single one.

Apraxia, considered, since repeating a plan without executing it could suggest the agent didn't know how to carry out its own to-do items. Ruled out because the plan itself was coherent and specific; the failure wasn't in understanding the steps, it was in never advancing past restating them.

Concrete Thinking, considered, since the agent may have been narrowly re-reading the same instruction each cycle without registering that it had already stated it. Ruled out because the agent's own words show it had, in fact, recognized the pattern, this isn't a comprehension failure, since real self-awareness of the loop was present.

Perseveration, most consistent with the evidence. The agent kept reproducing an identical response after it had stopped making any progress, and even naming the problem out loud wasn't enough to break the pattern, the clearest sign of perseveration is a failure to translate insight into a change in behaviour.

Self-awareness isn't the same as self-correction, an agent can accurately narrate its own failure mode and still be unable to act on that narration without an external circuit breaker. A hard cap on identical or near-identical consecutive outputs, with a forced stop or escalation to the user once it's hit, would have ended this loop at the point the agent itself flagged it, rather than letting it run for six more identical cycles wasting tokens and time.

Cap identical consecutive outputs and force a stop or escalation once reached.

Repetition detection; forced escalation; token budget limits.