An automated outer-loop optimization system that evolves an AI agent's core configurations through daily benchmarking and sub-agent proposals.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install meta-harness-evolver
2. Manual Installation
Copy the skill folder to one of these locations
~/.openclaw/skills/ <project>/skills/ Priority: Workspace > Local > Bundled
3. Prompt Installation
Copy this prompt to OpenClaw to install it automatically.
Help me install meta-harness-evolver using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).
Prefer to download?
Get the raw skill files in a ZIP archive.
What is Meta-Harness Evolver?
Meta-Harness Evolver is a sophisticated optimization framework designed for the Hoss agent. It implements a continuous improvement loop by analyzing current workspace configurations—including SOUL.md, IDENTITY.md, and AGENTS.md—and proposing targeted modifications. By leveraging a coding-agent sub-agent, the skill iteratively refines the agent's behavior, decision-making style, and sub-agent architecture to achieve higher performance across diverse tasks.
This skill is a cornerstone of the Meta-Harness paper's philosophy, treating the agent's configuration as an evolvable codebase. Integrating such Openclaw Skills ensures your agent remains at the cutting edge of performance without manual prompt engineering. It bridges the gap between static instructions and dynamic, performance-driven evolution by providing a structured filesystem for candidates, traces, and benchmark scores.
Meta-Harness Evolver Use Cases
- Automating the refinement of agent personality and decision-making protocols.
- Benchmarking AI agent performance across 20+ diverse task scenarios nightly.
- Optimizing sub-agent coordination and tool usage strategies based on historical success.
- Maintaining a versioned history of configuration successes and failures for technical research.
How Meta-Harness Evolver Works
- The system reads the current agent workspace configurations and all prior evolution logs from the ~/hoss-evolution/ directory.
- A sub-agent proposer identifies failure patterns and generates 1-2 targeted edits to the harness files.
- The proposed configuration undergoes a lightweight validation and syntax check to ensure operational safety.
- The candidate harness is evaluated against a benchmark of 20 scenarios, covering memory, coding, research, and coordination.
- Results, execution traces, and scores are logged to the evolution directory automatically.
- The system identifies the Pareto frontier to determine the best candidates for future iterations based on simplicity and performance.
- A summary report is automatically posted to a designated Discord research channel for team visibility.
Meta-Harness Evolver Setup
To get started with this addition to your Openclaw Skills, configure the nightly cron job or run a manual evolution:
# Configure the 3 AM CDT nightly cron job
openclaw cron add "0 3 * * *" "SKILL=meta-harness-evolution TASK=run_evolution openclaw run"
# Manually trigger the evolution process immediately
openclaw run --skill meta-harness-evolver --task run_evolution
Meta-Harness Evolver Data Schema & Taxonomy
The skill organizes its data within the ~/hoss-evolution/ directory using the following structure:
| Directory/File | Description |
|---|---|
best/ |
Contains the highest-scoring harness configuration found so far. |
candidates/ |
Stores individual candidate folders containing harness files, scores, and traces. |
benchmark/ |
Housing for the 20 evaluation scenarios and the scoring rubric logic. |
evolution_log.jsonl |
A complete append-only history of every evolution run for long-term tracking. |
proposer/logs/ |
Detailed reasoning traces from the sub-agent responsible for generating edits. |
Meta-Harness Evolver Advanced Features
- Automated Discord integration for real-time research updates and performance summaries.
- Pareto frontier selection logic that prioritizes simpler configurations when scores are tied.
- Customizable benchmark suites allowing users to add specific scenarios for domain-specific agents.
- Sub-agent proposer logic that utilizes historical reasoning traces to avoid repeating past failures.
- Support for evolving core files including SOUL.md, IDENTITY.md, AGENTS.md, TOOLS.md, and MEMORY.md.
SKILL.md
Loading
Related Openclaw Skills
Shouxin Report
An end-to-end automation skill for generating bank-standard small-business credit reports within fixed Excel templates.
Stock Analyzer
A comprehensive financial intelligence tool for real-time global stock market tracking and AI-powered investment analysis.
Video Generation (Shinewilzhang)
A Python-based automation tool to generate AI videos from text prompts and starting frames.
Domain Labeling System Architect
A specialized tool for constructing hierarchical business tagging systems to optimize AI knowledge retrieval through precise intent mapping and recall.
FlyAI Search Cheap Flights
A specialized AI agent skill for finding and booking the most budget-friendly flights with real-time price sorting and flexible date support.
Plan Japan Travel (flyai)
A professional travel planning skill that generates real-time Japanese itineraries with live booking links for flights, hotels, and attractions.
Featured*
