Agentless

Why we included this project

Agentless takes a different route from most coding agents: it never lets an LLM wander through a repository. Instead it splits the work into three explicit steps, first narrowing the bug down to the relevant files and functions, then generating a batch of candidate patches, and finally running regression and reproduction tests to pick the strongest fix. That structure makes the process easy to follow and cheap to run per issue, which appeals to researchers who want to study how each stage affects repair quality. If you're building automated debugging, triage, or CI systems that might suggest fixes, it's a useful reference to study. The repo ships with SWE-bench Lite and Verified configs, so you can compare your results against published baselines without building your own harness.