Awesome AI Agents › AI Developer Tools

OpenAutoCoder/live-swe-agent

⭐ 459 repository created 2025-11-13

Live-SWE-agent is a software engineering agent that revises and extends its own capabilities while working on a task, instead of relying on a fixed set of hand engineered tools. The authors argue that an agent is itself a software system and that a modern language model can already write and install the helper tools it needs, so the scaffold starts minimal and lets the model modify its own behaviour at runtime as it explores a repository and works toward a patch. The implementation is a thin layer over the mini-swe-agent framework: the repository ships a configuration file, and running mini with that configuration reproduces the setup, which keeps the amount of scaffold code that has to be trusted small. The repository also works as a benchmarking platform. Because one open scaffold is used for every model, scores become comparable across models, which the authors contrast with proprietary scaffolds that make published numbers hard to interpret. The leaderboard covers SWE-bench Verified and SWE-Bench Pro, and the README reports 79.2 percent on SWE-bench Verified with Claude Opus 4.5, 77.4 percent with Gemini 3 Pro, and 45.8 percent on SWE-Bench Pro. Complete artifacts from those runs, meaning agent trajectories, generated patches and result files, are attached to the 1.0.0 release and mirrored as Hugging Face datasets, so third parties can inspect the runs rather than trust summary numbers, and outside teams are invited to submit evaluations of their own models. The work is documented in an arXiv preprint by Xia, Wang, Yang, Wei and Zhang. It targets researchers comparing coding agents and models on equal terms, and engineers who want a compact self-extending scaffold to run against real issues in their own repositories.

https://github.com/OpenAutoCoder/live-swe-agent

agentllmself-evolvingsoftware-developmentswe-benchcoding-agentbenchmarkmini-swe-agentresearch