live-swe-agent

About

Live-SWE-agent rests on the observation that an agent is itself software, so an LLM-driven agent can extend and revise its own behavior while working — the implementation is deliberately a small config delta on mini-swe-agent rather than a new scaffold. Published results claim 79.2% on SWE-bench Verified with Claude Opus 4.5 and 45.8% on SWE-Bench Pro, self-reported through the project's own public leaderboard, which also serves as a platform for apples-to-apples model comparison. Trajectories, patches, and full run artifacts are published on Hugging Face for verification. Agent researchers use it to study runtime self-evolution and as a minimal baseline for fair model comparisons.