AI agents now have a way to quit while they're behind — before burning through compute on tasks they were never going to finish.
Researchers published a method that reads an LLM agent's internal hidden-layer activations to predict, as early as the first interaction round, whether an episode will fail. The system chains these predictions into an "abort cascade" — a series of calibrated gates, one per round, that kill a task when failure looks likely. On two models tested against the TextCraft benchmark, the cascade cut wasted inference compute by 47.1% for Qwen-2.5-7B and 37.2% for Llama-3.2-3B, while still letting genuinely successful runs through at user-set recall rates between 90% and 97%.
The key finding is where the signal comes from. Probes reading only observable agent behavior — what the agent actually does or says — saved roughly half as much compute and gained nothing from being combined with hidden-state probes. The internal representations already contain what behavior eventually reveals, just earlier. That matters because it means the fix doesn't require redesigning agents or changing how tasks are specified; it bolts on as an inference-time filter.
LLM agent compute costs are a growing pressure point for anyone running multi-step task pipelines at scale, and most current approaches wait for an agent to visibly fail before cutting losses. This work reframes that as an unnecessary tax — though the method was tested on a single game-like benchmark, and how well it transfers to messier real-world agent deployments remains an open question.