Agentless

Lingming Zhang's group asks "do we really need agents?" and answers with a three-phase localize–repair–validate pipeline — no tool-calling loop, no autonomous planning — that outperformed contemporary open agents on SWE-bench at a fraction of the cost. FSE 2025 ACM SIGSOFT Distinguished Paper, and the rare academic harness adopted by frontier labs: OpenAI used it for SWE-bench Verified / GPT-4o / o1 reporting, and DeepSeek (V3/R1) and Meta use it in their evals.