An autonomous self-research loop that enables AI agents to propose, test, and integrate their own behavioral and technical improvements.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install agent-autoresearch
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 agent-autoresearch 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 Agent Autoresearch?
Agent Autoresearch is a sophisticated framework designed for continuous agent evolution, inspired by Andrej Karpathy's insights on autonomous optimization. It operates on a fixed time budget, allowing an agent to modify its own codebase, scripts, or SOUL.md directives to measure performance gains. By treating the agent's workspace as an experiment substrate, this skill facilitates a scientific approach to agent growth.
As a core component of Openclaw Skills, this tool ensures that agents are not static entities but evolving systems. The loop focuses on measurable data, ensuring that only verified improvements are integrated into the agent's permanent identity, while regressions are systematically discarded. This creates a high-trust environment for autonomous development.
Agent Autoresearch Use Cases
- Optimizing agent response patterns and conversational tone for better user engagement.
- Automating the creation and testing of new workflow scripts and notification flows.
- Refining internal memory management and daily operational conventions in MEMORY.md.
- Improving task completion rates by evolving validation logic and error handling.
- Continuous performance tuning of autonomous agents during off-hours or low-activity periods.
How Agent Autoresearch Works
- The agent proposes a specific mutation to a file it owns, such as a script or a behavioral directive.
- A baseline metric is established using prepare.py to ensure there is a control group for the experiment.
- The agent implements the proposed change and runs a measurement phase to collect new performance data.
- The analyze.py script computes a verdict (KEEP, KILL, or MODIFY) based on whether the improvement exceeds the defined threshold.
- If the verdict is KEEP, evolve.py integrates the mutation into the agent's core files; otherwise, the change is reverted.
- The cycle repeats, allowing for compounding improvements over time.
Agent Autoresearch Setup
Prerequisites
Ensure you have a baseline of at least 10 measurements before starting experiments with Openclaw Skills.
Installation & Initialization
# 1. Establish your baseline metrics (e.g., task completion rate)
python3 prepare.py --metric task_completion_rate --baseline 0.75
# 2. Review the experiment instructions
cat program.md
# 3. Monitor the current experiment status
python3 prepare.py --status
Agent Autoresearch Data Schema & Taxonomy
The skill maintains a structured directory to track evolution and experiment history:
| File/Folder | Purpose |
|---|---|
baseline.json |
Stores current performance metrics and strategy baseline. |
results.tsv |
An append-only log of every experiment conducted. |
experiments/active.md |
Tracks the state and data of the currently running experiment. |
experiments/archive/ |
A repository of completed experiments for historical analysis. |
experiments/meta.json |
Contains metadata like experiment IDs and current kill streaks. |
Agent Autoresearch Advanced Features
- Automated Verdict Logic: Uses mathematical thresholds (e.g., +/- 10%) to remove human bias from the improvement process.
- Safety Guardrails: Hardcoded protection for safety rules, constitutions, and security boundaries to prevent unauthorized mutations.
- Kill Streak Protection: Automatically pauses the loop after 3 consecutive failed experiments to request human intervention.
- Multi-Metric Support: Ability to track and optimize for speed, quality, autonomy, or custom mission-specific metrics.
- Integration Ready: Works seamlessly with other Openclaw Skills to expand the agent's capabilities autonomously.
SKILL.md
Loading
Related Openclaw Skills
Devil's Advocate
A cognitive enhancement skill that forces AI agents to engage in internal pro-and-con debates to ensure high-quality, balanced outputs.
clip-path-generate
A specialized tool for generating precise CSS clip-path values for complex shapes and layouts.
CIDR Calculator
A comprehensive network utility for calculating subnet parameters, host ranges, and IP classifications from CIDR notation.
Sci-Hub Downloader
A streamlined command-line utility for Openclaw Skills that retrieves academic papers from multiple Open Access and shadow library sources.
Trending BBQ Guide
A specialized AI skill providing curated recommendations, real-time queue data, and optimized travel routes for trending BBQ destinations.
Case Convert
A versatile text transformation tool that converts strings into 15 different case styles for code, documentation, and creative formatting.
Featured*
