Autonomous research plugin for Claude Code — orchestrates four agents through evidence-backed improvement cycles
Agent
Use this agent to verify that an implemented code improvement actually works. It designs and runs ablation studies, benchmarks, and empirical tests to produce quantitative evidence of improvement or regression. It does NOT modify the implementation — it only observes and measures. Context: The implementer has made…
Agent
Use this agent to implement a specific, well-defined code improvement. It reads an improvement specification from a scout report, creates an implementation plan, makes minimal targeted code changes, runs tests, and captures the diff. Context: The research coordinator has a scout report with a top-ranked improvement.…
Agent
Use this agent when searching for state-of-the-art techniques, relevant research, and evidence-backed improvements applicable to a codebase. This agent specializes in scientific literature review, adjacent-field cross-pollination, and grounding improvement suggestions in published evidence. Context: The research…
Agent
Use this agent to review a completed research cycle, make an evidence-based accept/reject decision, update documentation, maintain the improvement backlog, and suggest the direction for future cycles. Context: All phases of a research cycle are complete and need final review. user: "Review cycle 3: scout report…