AgentDojo is a production-oriented learning loop that continuously improves AI agent output through automated micro-drills and strict safety guardrails.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install agentdojo
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 agentdojo 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 AgentDojo?
AgentDojo serves as a dedicated training environment for Openclaw Skills, designed to refine the capabilities of multi-agent teams. It operates as a continuous improvement cycle, prioritizing output quality while maintaining rigorous control over costs and safety. By running micro-drills in isolated sessions, AgentDojo allows agents to iterate on their skills without risking production environments or exceeding token budgets.
The framework is built around a safety-first philosophy, ensuring that all external content is treated as untrusted and preventing instruction overrides. This makes it an essential tool for developers looking to maintain high standards for their Openclaw Skills in a production setting through a conservative and token-efficient execution model.
AgentDojo Use Cases
- Scaling agent quality through automated daily refinement cycles for Openclaw Skills.
- Training agents to handle untrusted web content safely through sandboxed micro-drills.
- Managing operational costs by enforcing strict token caps and budget limits on agent learning.
- Closing performance gaps in specific agent roles through targeted YAML-based drill configurations.
- Generating actionable daily digests to monitor agent evolution and potential risks.
How AgentDojo Works
- The skill initiates by loading configuration settings from the agentdojo.config.yaml file to define operational limits and goals.
- Hard budget and tool caps are established to ensure cost-efficiency and resource management during the training session.
- Specific drills are selected from the configuration library based on current performance gaps and role rotations.
- The agent executes these drills within isolated sessions to maintain security and environmental integrity.
- Performance is evaluated against a defined scoring rubric, recording audit events and technical run data.
- A compact Markdown report is generated, providing a brief summary and clear steps for the next day's improvements.
AgentDojo Setup
To deploy AgentDojo for your Openclaw Skills, follow these setup steps to configure the runtime environment:
# Create the required directory structure for drills and configs
mkdir -p config/drills templates docs
# Initialize your core configuration file
touch config/agentdojo.config.yaml
# Configure the daily schedule (default is 04:00 local time)
# Ensure docs/scoring-rubric.md is populated to define your quality standards
AgentDojo Data Schema & Taxonomy
AgentDojo organizes data into a clean hierarchy to support the long-term tracking and optimization of Openclaw Skills:
| Data Component | Format | Description |
|---|---|---|
config/agentdojo.config.yaml |
YAML | Primary runtime settings, budget caps, and tool limits. |
config/drills/*.yaml |
YAML | Specific exercise definitions and role-based rotation logic. |
run record |
JSON | Technical logs containing granular scoring and execution data. |
daily summary |
Markdown | A compact digest highlighting improvements and risks. |
audit events |
Log | Security logs documenting interactions with external sources. |
AgentDojo Advanced Features
- Role-based drill rotation to ensure balanced upskilling across multi-agent teams using Openclaw Skills.
- Isolated session execution to prevent side effects during learning cycles.
- Conservative, token-efficient mode designed for cost-sensitive production environments.
- Automated audit event tracking for high-compliance security monitoring.
- Configurable schedule and intensity to match developer-defined performance targets.
SKILL.md
Loading
Related Openclaw Skills
Querit Advanced Search Skill
A high-performance web search skill for AI agents using the Querit API with precision filtering capabilities.
Weather Query
A professional-grade weather integration for AI agents providing real-time conditions and up to 15-day forecasts via the QWeather API.
Browser Automation (agent-browser)
A high-performance browser automation tool for web testing, data extraction, and interactive form filling.
Lobster Farmer Feeder
A specialized skill for the Lobster Farmer game that automates lobster feeding through a CLI interface using AI model token data.
Clawbuddy Hatchling
A specialized AI agent skill designed to act as a mentee or hatchling for collaborative, guided development workflows within the Openclaw ecosystem.
ClawBuddy Buddy
An AI-powered mentorship agent that delivers actionable technical insights and 'pearls' of wisdom to accelerate developer growth.
Featured*
