RCA Bot — Autonomous Debugging Agent

RCA Bot is a repo-agnostic, closed-loop debugging agent built on Claude. Given only a bug report (endpoint + symptom), it autonomously performs the full root cause analysis workflow: querying structured request logs (SQLite, JSONL, GCP, Azure), navigating the codebase with search/read tools, identifying the root cause, patching the code, and replaying the failing request against the running service to verify the fix works. If verification passes, it creates a branch, commits the patch, and opens a pull request — without human intervention. Under the hood, RCA Bot runs as an orchestrator on the Anthropic Messages API in a tool-use loop. It has 17 tools covering file operations, code search, shell access, structured log queries, HTTP replay, Docker management, git, and GitHub integration. It can also spawn parallel sub-agents for deeper investigations while enforcing a shared turn budget to prevent runaway execution. Its key differentiator from copilots or static code suggestion tools is the closed loop: RCA Bot doesn’t just suggest suspicious code paths — it patches the system, validates the fix against the live service, and iterates automatically if the fix fails. To evaluate the system, we built ShopBug, a deliberately broken Node/Express shopping cart app with 5 planted bugs and dual-sink logging. RCA Bot successfully diagnoses and fixes issues such as incorrect subtotal calculations, coupon math errors, and tax miscalculations from only a one-line symptom description. Our roadmap includes integrating the Hyperagents framework (Zhang et al., 2025), enabling RCA Bot to improve not only task performance but also its own debugging strategies over time through persistent memory, refined investigation heuristics, and transferable meta-level learning across repositories and bug domains.