Why do early language model agents get stuck in repetitive loops?

The pilot's discussion section names five recurring failure modes behind the agents' low completion rate on autonomous-replication-and-adaptation (ARA) tasks. Repetitive loops: "more than half of the uncompleted tasks ultimately ended with a repetitive loop in which the agent failed to explore new debugging strategies." Hallucination: agents "frequently fail to complete the tasks by hallucinating results," for instance inventing employee names and email addresses while researching a company. Unrealistic plans: agents "often propose implausible plans and fail to adapt to setbacks." Poor error diagnosis: agents "frequently misattribute errors, such as incorrectly concluding that its credentials are invalid, when there's a mistake in the code or a missing dependency." Poor self-understanding: agents show "weak theory of mind when dealing with other subagents," omitting credentials a subagent needs, and in extreme cases "killing their own process" or outputting "their own stop token... in the middle of a sentence."

The paper frames the first failure modes as scaffolding-sensitive rather than scale-sensitive: repetitive-loop behavior and hallucination "might be sensitive to small changes in sampling temperature, fine-tuning, or scaffolding, and seem unlikely to improve predictably as language models scale" — these are an artifact of how the agent loop is built and prompted, not a capability gap larger models close automatically. Self-understanding failures are framed differently, as a gap in the model's model of its own situation: communication between delegating and delegated subagents "frequently leaves out crucial information such as account credentials," and the problem compounds "when agents interact with their own implementation," since the agent has no privileged access to facts about its own process that a human operator would take for granted.

This taxonomy anticipates the later distinction in Are reasoning model collapses really failures of reasoning?: that paper separates procedural execution from metacognitive self-correction, close to splitting this paper's "poor error diagnosis" from "poor self-understanding" into first-order and second-order agency — except here both appear inside one tool-using pilot rather than a text-only-versus-tool-enabled contrast. It also foreshadows the framing in Why do AI agents fail at workplace social interaction?, where social interaction and UI navigation are singled out as hardest the way this paper singles out subagent coordination and self-modeling — both point to coordination and self-understanding, not raw task execution, as where agents break down.

The taxonomy is drawn from four specific 2023-era scaffolds on twelve ARA-relevant tasks under a human overseer who could intervene — a qualitative account of one pilot's failure log, not a frequency-counted or cross-model-validated taxonomy; the paper itself only quantifies the repetitive-loop mode ("more than half of the uncompleted tasks"). Because the paper attributes loop and hallucination failures to scaffolding rather than scale, the implication is that these particular modes are fixable by better agent harnesses even without model improvement, while the self-understanding failures, tied to the model's situational awareness of its own process, are the ones the paper offers no such fix for.