FailFast-RestartSmart: Predicting and Restarting SWE Agent Failures

A recent study presents FailFast-RestartSmart, a dual-phase controller aimed at enhancing the effectiveness of software engineering (SWE) agents that tackle repository-level challenges through extensive, context-building trajectories. This research, accessible on arXiv (2608.03222), highlights that prolonged failed runs often involve unnecessary exploration or looping, advocating for early detection of failures. FailFast operates as a lightweight 0.6B monitor, utilizing terminal and dense fail-to-pass supervision to foresee failures based on observable prefixes, eliminating the need for policy logits or hidden states. Upon failure detection, RestartSmart initiates a new rollout with the same policy, allowing the agent to review, apply, or disregard the interrupted repository diff. This strategy seeks to mitigate the risks of prematurely terminating successful trajectories while conserving resources on unsuccessful attempts. Results from SWE-bench demonstrate that this method can enhance efficiency without compromising success rates, contributing to the evolving domain of AI-driven software development, where autonomous coding agents are increasingly prevalent. The authors suggest a viable approach for failure prediction and restart, potentially lowering computational expenses and boosting the reliability of these agents.