claude-autoresearch

A plugin for Claude Code that runs autonomous, multi-milestone research loops. Inspired by Karpathy's autoresearch sketch, hardened into a real loadable plugin with verification at each step.

What it does

You hand it a research topic and a milestone structure. It plans, executes, and self-verifies — emitting a verification report per milestone (research/m{1,2,3}-*-verification.md) so each step is auditable before the next runs.

Status

• v1.0.0 shipped: 2026-04-16
• M3 smoke passed: 125× run on algo-trader-001. Algorithmic-trading research output cherry-picked back to main as c91651c.
• Skill fixes: 4 applied during M3 hardening.
• Distribution: plugin form. Load via claude --plugin-dir <path> or symlink to ~/.claude/plugins/autoresearch/.

Why I built it

The default Claude Code conversation loop is single-context — once you exhaust the window, you're done. I wanted a research workflow that survives context resets and self-evaluates as it goes, so a long-horizon task (multi-day research, deep technical investigation) can run without me babysitting it.

Connection points

• Direct lineage from Karpathy's autoresearch concept — same modify → run → measure → keep / revert loop, but with milestone-based verification and packaged as a Claude Code plugin.
• Pairs with agent-orchestrator (the daemon that spawns these loops) and research-orchestrator (the multi-Claude parallel-research design that informed it).
• The eval-platforms research report in this garden was produced by an earlier version of this pattern.

Repo is currently local-only. I'll publish once the milestone-template generator is split out.