SWE-Agent Autonomous Bug Fixing Implementation
SWE-Agent (Princeton NLP) is open-source agent for autonomous software development task solving. Unlike Devin, SWE-Agent is fully open-source, deployed on own infrastructure, requires no subscription to closed service.
How SWE-Agent Works
AgentComputer Interface (ACI) — specialized interface for agent interaction with codebase. Special commands: open, goto, search_dir, find_file, edit — optimized for code navigation. LLM-backbone: GPT-4o or Claude 3.5 Sonnet.
Work cycle: reads issue → explores codebase → forms hypothesis on cause → edits files → runs tests → iterates until passing.
Self-hosted Deployment
Docker container with Python environment. Sandbox based on Docker: isolated file system, restricted network access. Support for any LLM with OpenAI-compatible API.
Performance
On SWE-bench (benchmark on real GitHub Issues):
- GPT-4o backbone: ~38% resolution rate
- Claude 3.5 Sonnet backbone: ~43% resolution rate
- Best results for bug fixes with good tests
Implementation: 2–3 Weeks
Docker environment setup, GitHub workflow integration (GitHub Actions trigger), LLM-backend configuration, testing on representative issue sample from backlog.