Autoresearch skill
An autonomous iteration skill for agents: modify, verify, revert failed changes, and repeat the loop toward a target metric
- 6,834 ★ on the repository
- Last commit
- MIT
Broad access
Why this level
- Autonomously edits code in a loop, committing and reverting via git
- Runs shell commands to verify with tests and benchmarks
- Without guardrail hooks it may perform arbitrary actions in the project
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add skills/uditgoenka-autoresearchDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- Autonomously edits code in a loop, committing and reverting via git
- Runs shell commands to verify with tests and benchmarks
- Without guardrail hooks it may perform arbitrary actions in the project
This is third-party code. Review the repository files before installing.
What it does
Autoresearch turns an agent into a continuous improvement loop inspired by Karpathy's autoresearch. The user sets a goal and a mechanical metric, and the agent repeats: read the current state and git history, make one focused change, commit, run verification via tests or benchmarks, then keep the change if it improved things or revert it. Beyond the core loop there are separate commands: plan, debug, drive errors to zero, security audit via STRIDE and OWASP, edge-case generation, requirement probing, results analysis and a regression gate. In the orchestrator mode you state a goal in plain language and the agent selects the commands itself and runs the loop until the condition is met. For Claude Code there are guardrail hooks that block dangerous commands and reading of secrets.
Who it is for. For developers and quality engineers who want the agent to iteratively improve code or any measurable result on its own.
Good fit when
- You have a goal and a number you can measure: tests, a benchmark, a score
- You need to drive the error count to zero through repeated fixes
- You need an autonomous improvement loop with automatic rollback of failures
- You need a security audit or a regression gate before a release
Not a fit when
- There is no mechanical metric and quality can only be judged subjectively
- You cannot let the agent change and commit code autonomously
Example request
Run the improvement loop until all tests pass, and revert every change that makes things worseLimitations
The loop only works where a verifiable metric and a way to run it exist; it does not fit subjective tasks. The guardrail hooks are Claude Code only, while OpenCode and Codex get the core skill without them, so protection there must be handled separately. The agent edits code, commits and reverts on its own, so run it on a repository with a working git.
How to disable. Remove the autoresearch plugin via /plugin or delete the skill and command folders from the agent's directory. Individual hooks can be turned off with AR_DISABLE_* environment variables.
FAQ
What do I set before running?
A goal, a mechanical metric and the edit scope. The agent runs a one-time setup, takes a baseline, and only then starts the loop.
Do the safety hooks work everywhere?
No, the full hook set is Claude Code only. OpenCode and Codex get the core skill and commands without hooks.
Discussion
No questions or reviews yet.
Related
SkillsOrchestrator skill: from an idea the agent writes requirements and a spec, splits work into tasks, builds the project, checks it
Over twenty skills for coding agents: MCP and hooks, plan mode, a bug-fix protocol, prompting and a bash safety hook
A library of 30-plus instruction skills for coding agents: planning, building, testing, hallucination guards and shipping
A Claude Code skill that drives a hard task through explore, plan, execute and reflect, writing every decision to disk.
14 skills that find the gaps in a task before the code: interviews, plans, assumption tests and proof that a bug fix works
A skill that runs a full refactor so agents read code more easily: it finds hidden links and needless dependencies.
README in short
The README describes Autoresearch as an autonomous improvement loop for Claude Code, OpenCode and Codex that generalizes Karpathy's approach to any domain with a measurable metric. The core loop makes one change per iteration, commits before verifying, runs a mechanical check and reverts anything that did not improve the result, with git serving as experiment memory. Beyond iteration there are fourteen commands: plan, debug, fix, security, ship, scenarios, requirement probing and others, while an orchestrator assembles the right pipeline for a plain-language goal. It documents the Claude Code guardrail hooks that block dangerous commands, reading secrets and context bloat. Installation is via npx skills, a plugin or manual copying, and the architecture is split into a thin routing file plus separate commands to save context.
README badge
Original SKILL.md
--- name: autoresearch description: "Autonomous iteration loop: modify, verify, keep/discard against any metric" version: 2.2.2 --- # Autoresearch — Autonomous Goal-directed Iteration ## Safety Invariants (all subcommands) LOOP (N iterations or until done): 1. Review current state + git history + results log 2. Pick the next change (based on what worked, what failed, what's untried) 3. Make ONE focused change 4. Git commit (before verification) 5. Run mechanical verification (tests, benchmarks, scores) 6. If improved keep. If worse git revert. If crashed fix or skip. 7. Log the result 8. Repeat until N iterations complete or goal is met.