Autoagent

Automatically improve agent guidance through iterative testing and scoring.

Autoagent is an OpenClaw skill that optimizes any agent guidance (prompts, AGENTS.md entries, skill definitions) using a Karpathy-style training loop with cron-triggered optimization.

What It Does

1. Setup - Asks where your guidance lives and what it should do
2. Sandbox - Copies guidance to a test folder with fixtures
3. Optimize - Runs every 5 minutes via cron:
   • Analyzes current guidance
   • Proposes improvement
   • Tests with subagent
   • Scores result
   • Keeps or discards change
4. Logs - Check scores.md for history

Quick Start

/autoagent

How It Works

Setup Phase

Every invocation starts fresh with interactive setup:

1. Sandbox Location - Where should I create the folder? (absolute path)
2. Success Criteria - Discussion to define what "good" looks like, then propose draft scoring.md for approval
3. External Scripts - Any scripts/tools the guidance relies on?
4. Cron Schedule - How often to run (default: 5 minutes)

Then creates a sandbox at your specified path with:

• guidance-under-test.md - Original (read-only)
• current-guidance.md - Working version
• fixtures/test-cases.json - Test inputs
• scoring.md - User-approved success criteria
• scores.md - Score history
• scripts/ - (optional) copies of referenced scripts

Iteration Loop

Each cron run:

1. Analyze current guidance + score history
2. Propose ONE specific edit
3. Apply edit to current-guidance.md
4. Run subagent with new guidance
5. Score result (0-100)
6. Log decision (keep/discard)
7. Revert if score declined
8. Check for plateau (auto-pause after 10 identical scores)

Example

User: /autoagent

Agent: Where should I create the sandbox folder? Default: `../../autoagent-sandbox/` (resolves to `/clawd/autoagent-sandbox/`)
> news

Agent: Let's define how we'll measure success. What does a "good" result look like for this task?
> It should summarize news into morning briefs with bullet points

Agent: What format should the output be in?
> Markdown with sections for Top Stories, Industry News, and Quick Hits

Agent: [proposes draft scoring.md]

[User approves scoring]

Agent: Does the guidance rely on any scripts, tools, or external software?
> No

Agent: Run optimization every 5 minutes (default), or different interval?
> Default is fine

[Sandbox created at `../../autoagent-news/` → `/clawd/autoagent-news/`, cron started]

Agent: Optimization started at `/clawd/autoagent-news/`. I'll check back every 5 minutes. Monitor progress in `scores.md`.

[5 minutes later...]
- Score: 45 → 52 (kept change)
- Score: 52 → 48 (discarded)
- Score: 48 → 61 (kept change)
...

Files

File	Description
SKILL.md	Full skill definition
setup-prompt.md	Setup phase questions
iteration-prompt.md	Iteration loop instructions
templates/	Score and fixture templates
examples/	Test examples

Requirements

• OpenClaw with cron support
• LLM for guidance analysis
• File system access

License

MIT