'Coding Agents Think Ahead of Time': Latent Probes Predict Outcomes 25 Steps Early

arXiv / Hacker News·medium signal

A July 6 arXiv paper by André Silva, Han Tu, and Martin Monperrus shows logistic-regression probes on a coding agent's hidden states can decode whether code parses and passes tests at AUC up to 0.83 — and that these representations 'run ahead' of the agent's own edits, predicting outcomes up to 25 steps in advance (a 'latent programming horizon'). Probes transferred across two models and two benchmarks without retraining. The interpretability result suggests agents internally 'know' where their edits are heading before they make them.