Hidden-State Probes Predict LLM Agent Failure Early, Saving Up to 47% Inference Compute

Researchers show that failure in LLM agent episodes is predictable early from the agent's internal representations. Lightweight per-round probes on hidden activations anticipate eventual episode failure as early as the first interaction round, where scorers reading only observable behavior are barely better than chance. The signal is turned into a practical abort cascade: one distribution-free calibrated gate per round, with per-round recall budgets jointly searched so that eventually-successful episodes survive all gates at a user-specified global rate. Across two agent models on TextCraft, the cascade meets every recall target from 90% to 97% and, at the 90% target, saves 47.1% ± 10.3% (Qwen-2.5-7B) and 37.2% ± 8.8% (Llama-3.2-3B) of inference compute, 1.6-1.7x the best single-gate policy. An identical cascade reading only behavior saves roughly half as much, and adding behavioral features to the probe yields no further gain: hidden states capture what behavior reveals. The work also characterizes the sample complexity of certifying high recall targets, telling practitioners which recall promises their data can and cannot back.