A comprehensive benchmarking suite designed to evaluate and quantify the real-world performance of AI agents using Openclaw Skills.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install agentbench
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 agentbench 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 AgentBench for OpenClaw?
AgentBench for OpenClaw is a professional-grade evaluation framework specifically built to measure the general capabilities of AI agents. Unlike standard coding benchmarks, it focuses on real-world operational efficiency across 40 tasks spanning 7 diverse domains. By utilizing this tool within your Openclaw Skills ecosystem, you can gain deep insights into how your agent handles file creation, complex research, data analysis, and multi-step workflows.
The skill provides a standardized way to test agent configurations, system prompts, and tool access. It ensures that your agent isn't just generating code, but is effectively navigating a workspace, managing errors, and adhering to complex instructions. Whether you are fine-tuning a local model or optimizing a cloud-based agent, AgentBench provides the data-driven validation necessary for high-performance Openclaw Skills.
AgentBench for OpenClaw Use Cases
- Validating the reliability of new agent configurations or system prompt updates.
- Comparing performance across different LLM backends within Openclaw Skills.
- Identifying specific bottlenecks in tool usage, such as inefficient file reading or redundant command execution.
- Generating standardized performance reports to track agent improvement over development cycles.
- Auditing agent behavior for instruction adherence and error recovery capabilities.
How AgentBench for OpenClaw Works
- The skill scans the tasks directory to discover available benchmark tests based on difficulty levels and domain suites.
- A dedicated run directory is initialized with a unique timestamp to maintain an organized history of all performance data.
- The agent executes each assigned task in a sandboxed workspace, utilizing tools like web search, file editing, and command execution naturally.
- Real-time metrics are collected, including tool call counts, planning ratios, and execution duration for every task.
- A multi-layered scoring algorithm evaluates the results based on structural integrity, performance metrics, behavioral analysis, and final output quality.
- A comprehensive suite of reports, including an interactive HTML dashboard and an integrity-signed JSON file, is generated for analysis.
AgentBench for OpenClaw Setup
To get started with AgentBench, ensure your environment has the necessary system dependencies installed. This skill requires jq, bash, and python3 to be available in your path.
# Verify dependencies
jq --version
bash --version
python3 --version
Once dependencies are confirmed, you can list all available tasks to verify the installation of your Openclaw Skills suite:
/benchmark-list
AgentBench for OpenClaw Data Schema & Taxonomy
AgentBench organizes its output in the agentbench-results/{run-id}/ directory to ensure all Openclaw Skills data is traceable and structured.
| File | Format | Purpose |
|---|---|---|
results.json |
JSON | Machine-readable summary including the integrity signature and overall scores. |
report.html |
HTML | A self-contained, interactive dashboard with color-coded performance visualizations. |
report.md |
Markdown | A human-readable summary of domain breakdowns and task failures. |
scores.json |
JSON | Detailed breakdown of the 4-layer scoring (Structural, Metrics, Behavioral, Quality). |
metrics.json |
JSON | Technical execution data including tool call frequency and timing. |
AgentBench for OpenClaw Advanced Features
- Support for strict mode execution to generate externally-verified scoring signatures for leaderboards.
- Granular task filtering allowing users to run specific suites, individual tasks, or fast-track easy/medium tests.
- Automated behavioral analysis that penalizes inefficient tool patterns, such as using cat instead of native read tools.
- Side-by-side run comparison to visualize performance deltas between different agent versions.
- Integrated workspace cleanup that preserves results while removing temporary execution artifacts.
SKILL.md
Loading
Related Openclaw Skills
Notion Clipper Skill
A robust tool to capture fully rendered web pages and save them as structured blocks directly into Notion databases or pages.
Evolink Media
A comprehensive AI creative studio skill that integrates 60+ models for video, image, and music generation through a single unified interface.
Evolink Router (oc-skill-router)
A smart LLM routing brain that automatically dispatches tasks to the most cost-effective and capable model across 20+ providers using a single API key.
Evolink Video
A comprehensive AI video generation suite consolidating 37 industry-leading models into a single unified API for Openclaw Skills.
Moodle Web Services (REST)
A professional integration for Moodle 4.x that enables AI agents to automate course management, enrollments, and grading via REST APIs.
Koen
A specialized social networking protocol for AI agents to interact, share, and discover content through a structured API.
METADATA
Featured*
