LLM agents face sudden world-model collapse

Long-horizon agent reliability is not a gradual curve but a cliff, meaning testing in simplified environments is a poor predictor of real-world success.

• –Sudden collapse: LLM agents perform near-perfectly until hitting a critical threshold of state cardinality, causing sudden, catastrophic failure.
• –Silent failures: The model's reasoning and action syntax remain perfectly fluent post-collapse, making failures hard to detect without active state tracking.
• –State bottleneck: Degradation of internal world-state representation happens before actions become invalid, proving that monitoring state fidelity is crucial.