What makes agent evaluation different from traditional software testing?

AI agents are non-deterministic and may take multiple valid paths to reach a goal. This skill focuses on outcome-based evaluation and weighted rubrics rather than checking for specific, rigid execution steps.

What is the '95% Finding' in agent performance?

Research indicates that token usage, the number of tool calls, and the specific model choice explain 95% of performance variance. This skill helps you optimize these specific drivers for better results.

Can this skill help with production monitoring?

Yes, it includes patterns for sampling real usage data, running it through evaluation pipelines, and setting up dashboards to track quality trends and detect performance cliffs over time.

How do I implement LLM-as-judge using this skill?

The skill provides guidance on designing evaluation prompts that use a high-reasoning model to assess agent outputs against specific criteria like factual accuracy, citation quality, and process efficiency.

AI Agent System Evaluation

Name: AI Agent System Evaluation
Author: ngxtm

byngxtm

•

データサイエンスとML

Builds robust evaluation frameworks to measure, validate, and optimize AI agent performance and context engineering strategies.

This skill provides a comprehensive framework for evaluating complex AI agent systems, moving beyond simple string matching to outcome-focused assessment. It enables developers to implement multi-dimensional rubrics—covering factual accuracy, completeness, and tool efficiency—while utilizing LLM-as-judge patterns and human-in-the-loop validation. By systematizing test set design and complexity stratification, it helps teams identify performance regressions, optimize token budgets, and validate context engineering choices to ensure reliable agentic behavior in production environments.

主な機能

01Multi-dimensional rubric design for accuracy, completeness, and tool efficiency

02LLM-as-judge implementation for scalable automated assessment

03Performance variance analysis based on token usage and model choice

04Complexity-stratified test set generation from simple to multi-step reasoning

05Continuous evaluation pipelines for regression detection in CI/CD

065 GitHub stars

ユースケース

01Measuring the impact of context degradation on agent performance and accuracy

02Creating automated quality gates to prevent agent logic regressions before deployment

03Benchmarking different LLM models or context strategies for a specific agentic task

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add ngxtm/devkit evaluation

For use in Claude.ai and ChatGPT

Download Skill