Two steps instead of an agent loop
On 1 July 2024 the University of Illinois showed Agentless: a fixed two-phase pipeline — locate, then repair — with no autonomous tool calls. On SWE-bench Lite it reached 27.33% at $0.34 per issue, the best and the cheapest among open-source approaches.
Why it matters
Until then the complexity of an agent was taken to be what produced the gain. The work set a cheap loop-free pipeline beside it and showed that among open systems the pipeline wins — so part of the gain came not from autonomy but from what surrounded it.
The pipeline: hierarchical localisation from file to class or function and then to line, followed by generating repairs and picking the best by tests. The model is GPT-4o (gpt-4o-2024-05-13). The side result turned out to matter more than the main one. The authors went through SWE-bench Lite by hand and found that 4.3% of issues carry the exact ground truth patch in the description itself, a further 10.0% describe the exact steps to the solution, 9.3% lack information needed to solve them, and 4.3% point down a misleading path. From what remained they assembled a stricter set, SWE-bench Lite-S. What the record does not claim. The 27.33% is the best only among open-source approaches, not in general: the paper's own table places Alibaba Lingma Agent (33.00%), Factory Code Droid (31.33%), AutoCodeRover-v2 (30.67%) and CodeR (28.33%) above it. The claim that 85% of autonomous agent failures come from over-modifying files does not appear in the paper.