SetupX: LLM agents learn from past failures to fix code repos with 92% success

Setting up a code repository to run its documented features correctly is notoriously difficult, even for advanced LLM agents. Dependencies clash, toolchains are missing, and installations fail in ways that are hard to diagnose. Existing agents struggle to transfer knowledge across repositories, handle irreversible state changes during multi-step fixes, and reliably verify that a setup is truly functional. To tackle this, a team of researchers introduced SetupX, a framework that learns from past failures and systematically resolves these issues.

SetupX is built on three key innovations. First, it constructs a Self-Evolving Experience Representation (XPU) that captures setup signals, textual guidance, and executable actions, allowing the agent to dynamically transfer verified fixes to unseen repositories. Second, it uses Experience-Augmented Speculative Execution backed by a LIFO Docker snapshot stack, so the agent can safely trial fixes and roll back to known-good states without corrupting the environment. Third, a Prosecutor-Judge Verification Protocol separates evidence collection from final judgment, enabling more reliable verification beyond superficial build-time metrics. In evaluation, SetupX achieved a 92% pass rate on carefully crafted benchmarks, outperforming the strongest baseline by over 19%. It excels particularly in complex multi-repository setups that require coordinating multiple interconnected services across different containers.