Agentless: less agent complexity, more results on SWE-bench
In one sentence UIUC publishes Agentless: a two-phase pipeline (localize fault, generate repair) without complex agent loops. Outperforms AutoCodeRover and SWE-agent on SWE-bench. Top open submission on SWE-bench at publication time. Challenges the assumption that more agent complexity equals better results.
In 2024 everyone is building increasingly complex agents to fix software bugs: reasoning loops, multiple tools, self-criticism, recursive planning. Agentless does the opposite and works better.
The idea is simple: given a bug report (a GitHub issue), first locate where the problem is in the code (which files, which functions), then generate the patch to fix it. Two phases, zero autonomous loops, no self-modifying agent.
The result: it beats much more complicated systems on SWE-bench, the standard benchmark for agents that solve bugs on real Python repositories. At the time of publication it is at the top of the open-source leaderboard.
The lesson is important: agentic complexity can be a problem, not a solution. A well-designed pipeline with clear phases can outperform an agent that "thinks freely". This finding will influence how subsequent systems are designed.
Companies
UIUC
Tools
—
Tags
Sources