Agentic Evaluation provides robust patterns and techniques for building AI agents that evaluate, critique, and refine their own outputs through iterative feedback loops.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install agentic-eval
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 agentic-eval 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 Agentic Evaluation Patterns?
The agentic-eval skill is a sophisticated framework within the Openclaw Skills ecosystem designed to move beyond single-shot AI generation. It introduces systematic patterns for self-improvement, allowing agents to assess their work against specific rubrics and optimize results until they meet high-quality standards.
By implementing these patterns, developers can create more reliable AI systems that handle quality-critical tasks like complex coding, analytical reporting, and compliance-driven content creation. This skill bridges the gap between raw LLM output and production-ready results by embedding evaluation directly into the agentic workflow.
Agentic Evaluation Patterns Use Cases
- Implementing self-critique and reflection loops for high-stakes content generation.
- Building evaluator-optimizer pipelines to separate generation logic from quality control.
- Creating test-driven code refinement workflows that automatically fix bugs via unit test feedback.
- Designing rubric-based or LLM-as-judge systems for objective output ranking.
- Automating iterative improvement for technical reports, analysis, and structured data outputs.
How Agentic Evaluation Patterns Works
- Generation: The agent creates an initial draft or solution based on the user prompt.
- Evaluation: A secondary process (or the same agent) assesses the output against defined success criteria or rubrics.
- Critique: The system generates structured feedback (typically in JSON) identifying specific areas of failure or improvement.
- Refinement: The agent processes the critique and regenerates the output to address the identified issues.
- Convergence: The loop repeats until the evaluation score meets the threshold or the maximum iteration limit is reached.
Agentic Evaluation Patterns Setup
To integrate agentic-eval into your project, ensure your environment is configured for Openclaw Skills and follow these steps:
# Install the agentic-eval skill patterns
openclaw install agentic-eval
# Configure your LLM judge credentials
export EVAL_JUDGE_MODEL="gpt-4o"
Ensure you have a structured logging system in place to track the trajectory of refinements during the iteration loops.
Agentic Evaluation Patterns Data Schema & Taxonomy
The skill utilizes structured JSON for evaluation feedback and maintains an iteration history to ensure convergence.
| Data Component | Format | Description |
|---|---|---|
| Critique Results | JSON | Maps criteria to PASS/FAIL status and specific feedback strings. |
| Evaluation Rubric | Dictionary | Defines weighted dimensions (e.g., accuracy, clarity) for scoring. |
| Refinement Trajectory | List | A log of all previous iterations, scores, and critiques for debugging. |
| Convergence Status | Boolean | Indicates if the output has met the required quality threshold. |
Agentic Evaluation Patterns Advanced Features
- Multi-Agent Orchestration: Separate 'Generator' and 'Evaluator' agents to prevent self-bias in quality assessments.
- Test-Driven Refinement: Automatic integration with pytest to drive code fixes based on real-time execution errors.
- Weighted Rubrics: Customizable scoring systems that prioritize specific dimensions like safety or technical accuracy.
- Convergence Detection: Intelligent early-stopping mechanisms that terminate the loop if quality improvements plateau.
SKILL.md
Loading
Related Openclaw Skills
AI Text Detector
An intelligent linguistic analyzer that estimates the probability of text being AI-generated based on structural and stylistic features.
mergeiq
A provider-agnostic complexity scoring engine that quantifies the depth of code changes using a multi-dimensional framework.
Resume Screening Agent Skill
A professional AI-driven resume evaluation tool that uses a funnel-based methodology to identify the most suitable and authentic candidates for any role.
Dev Progress Governor
An execution-governance skill that enforces commit discipline, generates messages, and maintains structured progress logs.
Wallet Overview
A powerful tool for monitoring multi-chain wallet balances and activity history through a single interface.
DreamAct Face Animation
A high-fidelity AI animation skill that transfers expressions and head poses from a source video to static face images.
Featured*
