autoloop

name	autoloop
description	Autonomous single-metric optimization loop that edits a target file, runs experiments, commits improvements, and reverts regressions until a numeric goal is reached.
compatibility	Created for Zo Computer
metadata	{"author":"marlandoj.zo.computer","origin":"karpathy/autoresearch patterns","version":"1.0.0"}

Autoloop — Autonomous Optimization Skill

Quick Start

# 1. Create a program.md in your project directory (see template)
cp Skills/autoloop/assets/template.program.md my-project/program.md
# Edit program.md with your metric, target file, and run command

# 2. Run the loop
bun Skills/autoloop/scripts/autoloop.ts --program my-project/program.md

# 3. Stop anytime with Ctrl+C — progress is saved via git commits

---

MCP Server

The autoloop skill exposes an MCP server for AI assistants to start, monitor, and manage optimization campaigns.

Available Tools

Tool	Description
autoloop_start	Start an optimization campaign with a program.md file
autoloop_status	Check the status of a running or completed campaign
autoloop_results	Get results.tsv data from a completed campaign
autoloop_stop	Gracefully stop a running campaign
autoloop_list	List recent autoloop campaigns

Usage

Stdio transport (for Claude Desktop, Cursor):

bun /home/workspace/Skills/autoloop/scripts/mcp-server.ts

HTTP transport (for network access):

# Start the server
bun /home/workspace/Skills/autoloop/scripts/mcp-server-http.ts

# Or run as a Zo service (recommended)
# Service URL: https://autoloop-mcp-marlandoj.zocomputer.io

Example: autoloop_start

{
  "program": "/home/workspace/my-project/program.md",
  "executor": "claude-code",
  "dryRun": false
}

Example: autoloop_status

{
  "program": "/home/workspace/my-project/program.md"
}

How It Works

1. Reads your program.md for configuration
2. Creates a git branch autoloop/{name}-{date}
3. Runs the baseline (first experiment, no changes)
4. Loops forever:
   • Agent proposes a change to the target file
   • Commits the change
   • Runs the experiment via your run command
   • Extracts the metric via your extract command
   • If improved → keep commit, advance branch
   • If regressed → git reset --hard HEAD~1
   • If crashed → log, attempt fix (max 3), skip if unfixable
5. Stagnation detection triggers increasingly radical exploration
6. Writes results.tsv with full experiment history

Options

--program <path>     Path to program.md (required)
--executor <name>    Executor to use for proposals (default: claude-code)
--resume             Resume from existing autoloop branch instead of creating new
--dry-run            Parse program.md and show config without running

Scheduling as a Zo Agent

The autoloop can run overnight as a scheduled agent:

Create a Zo agent that runs:
bun /home/workspace/Skills/autoloop/scripts/autoloop.ts --program /home/workspace/my-project/program.md

Cost guardrails (maxExperiments, maxDurationHours, maxCostUSD) prevent runaway spending.

References

• assets/template.program.md — Template for new programs
• references/autoresearch-patterns.md — Design notes from karpathy/autoresearch
• Spec: Skills/zo-swarm-orchestrator/specs/autoresearch-patterns-spec.md (Spec 3)