Code Guardian

Code Guardian is a fully autonomous five-agent AI pipeline that scans any GitHub repository for bugs and security vulnerabilities, patches every issue it finds, writes targeted pytest tests, runs them in an isolated sandbox, and opens a ready-to-merge Pull Request — all without human intervention. The pipeline runs on AMD Instinct MI300X via AMD Developer Cloud, using the ROCm open-source GPU stack and vLLM to serve Llama 3.3 70B. The five agents are: Scanner (pyflakes + LLM scan on the highest-risk files), Fixer (generates complete patched files using Python stdlib only), Test Writer (one verification test per bug using string-search inspection), Executor (runs tests in an isolated sandbox), and Reporter (opens a GitHub PR with diffs and AI-written description). The key innovation is the intelligent retry loop. When tests fail, the Executor asks the LLM to classify why — patch_wrong routes back to the Fixer, test_wrong routes back to the Test Writer, and environment errors surface immediately. This prevents infinite loops and ensures the right agent is retried. Max 3 retries per failure type. On our demo run against a purpose-built vulnerable repository, the system found 12 bugs including SQL injection, hardcoded secrets, MD5 password hashing, path traversal, and connection leaks — patched all of them, wrote 12 tests, and opened a pull request in under 2 minutes.