에이전트 스킬 성능을 A/B 테스트하는 평가 프레임워크

agent-skills-eval A test runner for Agent Skills . Write a SKILL.md , drop in some evals, and find out — empirically — whether your skill actually makes the model better at the task. Documentation · Quickstart · SDK · agentskills.io Why this exists Agent Skills — the open standard from Anthropic for giving agents domain knowledge — make it easy to ship a SKILL.md and assume your agent is now better at the task. The hard part is proving it. agent-skills-eval is the missing piece. It runs your skill against the same prompts twice — once with_skill loaded into context, once without_skill (baseline) — has a judge model grade both outputs, and gives you a side-by-side report. If the skill doesn't make a measurable difference, you'll see it. If it does, you have receipts. It's the test framework for the Agent Skills ecosystem, separated from any specific agent runtime so it works wherever your skills do. Quickstart npx agent-skills-eval ./skills \ --target gpt-4o-mini \ --judge gpt-4o-mini \ --baseline \ --strict That's it. Point it at a folder of skills, give it a target model and a judge model, and it produces a workspace with full artifacts and a static HTML report. agent-skills-workspace/ └── iteration-1/ ├── meta.json # run metadata ├── benchmark.json # rolled-up pass/fail per skill ├── eval-basic/ │ ├── with_skill/ # output, timing, judge grading │ └── without_skill/ # ↑ same, with the skill stripped └── report/ └── index.html # the visual report Open iteration-1/report/index.html and you have a real, evidence-backed answer to "is my skill working?" What you get with_skill vs without_skill Every eval runs both ways so you can see the actual lift from the skill — or its absence. Judge-graded outputs Use any chat model as a judge. Pass/fail with cited assertions, not vibes. TypeScript SDK + CLI One-liner CLI for CI, full SDK for custom pipelines, custom providers, and dashboards. OpenAI-compatible by default Works out of the box with OpenAI, Together, Groq, Anthropic via OpenAI-compat layers, local Llama servers — anything that speaks the OpenAI chat API. Tool-call assertions Deterministic checks for agents that call tools, not just generate text. Portable artifacts JSON + JSONL all the way down. Run today, diff tomorrow. Plug into your own dashboard. Static HTML reports A drop-in report site you can publish anywhere — no infrastructure. Fully spec-compliant Implements the full agentskills.io specification : SKILL.md validation, evals/evals.json , official iteration-N artifact layout, frontmatter rules. Install npm install agent-skills-eval Or run directly without installing: npx agent-skills-eval --help How it works The mental model is straightforward. For every eval defined in your skill: ┌─────────────────────────────┐ │ same prompt │ └───────────────┬─────────────┘ │ ┌───────────────┴─────────────┐ ▼ ▼ ┌──────────────┐ ┌──────────────┐ │ with_skill │ │without_skill │ │ SKILL.md in │ │ baseline, │ │ context │ │ no skill │ └──────┬───────┘ └──────┬───────┘ │ │ ▼ ▼ target model target model │ │ ▼ ▼ output output │ │ └──────────┬──────────────────┘ ▼ ┌─────────────┐ │ judge │ scores both against │ model │ the same assertions └──────┬──────┘ ▼ pass / fail per side The judge sees the eval's expected_output and assertions and grades each side independently. The --baseline flag is what enables the comparison; without it you only get the with_skill run. YAML config For anything beyond a quick command, drop a config file at the root of your project: # agent-skills-eval.yaml root : ./skills workspace : ./agent-skills-workspace baseline : true target : gpt-4o-mini judge : gpt-4o-mini baseUrl : https://api.openai.com/v1 apiKeyEnv : OPENAI_API_KEY include : - " skills/** " exclude : - " **/draft-* " concurrency : 4 layout : iteration strict : true report : enabled : true title : Agent Skills Report logging : format : pretty # pretty | jsonl | silent verbose : false color : auto targetParams : temperature : 0 judgeParams : temperature : 0 OPENAI_API_KEY=... npx agent-skills-eval --config agent-skills-eval.yaml CLI flags always override config values. SDK For programmatic use — CI pipelines, custom dashboards, multi-skill rollups — drive the evaluator from TypeScript: import { OpenAICompatibleProvider , consoleReporter , evaluateSkills , } from "agent-skills-eval" ; const provider = new OpenAICompatibleProvider ( { baseUrl : "https://api.openai.com/v1" , apiKey : process . env . OPENAI_API_KEY ! , model : "gpt-4o-mini" , providerName : "openai" , } ) ; const result = await evaluateSkills ( { root : "./skills" , workspace : "./agent-skills-workspace" , baseline : true , concurrency : 4 , workspaceLayout : "iteration" , strict : true , target : { model : provider . model , provider } , judge : { model : provider . model , provider } , onEvent : consoleReporter ( ) , } ) ; console . log ( result ) ; Stream events to a file as JSONL for downstream analysis: import { jsonlReporter } from "agent-skills-eval" ; const reporter = jsonlReporter ( { file : "./events.jsonl" } ) ; await evaluateSkills ( { /* ... */ onEvent : reporter . onEvent } ) ; await reporter . close ( ) ; Load YAML config programmatically: import { loadConfigFile } from "agent-skills-eval" ; const config = loadConfigFile ( "./agent-skills-eval.yaml" ) ; Custom providers Bring any backend by implementing the Provider interface — five fields, one method: import type { Provider , ProviderResult } from "agent-skills-eval" ; export const provider : Provider = { name : "my-provider" , model : "my-model" , async complete ( prompt : string ) : Promise < ProviderResult > { return { provider : "my-provider" , model : "my-model" , output : "model output" , latencyMs : 0 , inputTokens : 0 , outputTokens : 0 , costUsd : 0 , } ; } , } ; Useful for: local model servers (Ollama, vLLM, llama.cpp), proprietary internal APIs, mock providers in unit tests, or routing layers in front of multiple providers. Skill layout A skill is a folder. The minimum is a SKILL.md . Add evals/evals.json and you can evaluate it. my-skill/ ├── SKILL.md ├── references/ │ └── notes.md ├── scripts/ │ └── helper.sh └── evals/ ├── evals.json └── files/ └── input.csv SKILL.md : --- name : my-skill description : Analyze small CSV files. license : MIT compatibility : Works with text-capable chat models. --- When given a CSV file, identify the most important trend and cite the relevant rows. evals/evals.json : { "skill_name" : " my-skill " , "evals" : [ { "id" : " basic " , "name" : " basic behavior " , "prompt" : " Use the attached data to summarize revenue. " , "files" : [ " evals/files/input.csv " ], "expected_output" : " The response identifies the highest revenue month. " , "assertions" : [ " The output identifies the highest revenue month. " ] } ] } If you skip assertions but provide expected_output , the SDK promotes the expected output into a judge assertion automatically — so a minimal agentskills.io eval file produces meaningful pass/fail grading without extra work. CLI options npx agent-skills-eval [root] \ --config agent-skills-eval.yaml \ --workspace ./agent-skills-workspace \ --baseline \ --target gpt-4o-mini \ --judge gpt-4o-mini \ --base-url https://api.openai.com/v1 \ --api-key-env OPENAI_API_KEY \ --include " skills/** " \ --exclude " **/draft-* " \ --concurrency 4 \ --layout iteration \ --strict \ --log-format pretty \ --report Logging modes : pretty for humans, jsonl for machines, silent for quiet CI. Reports The static HTML report is built from disk artifacts and shows everything you'd want for skill iteration: Pass rate by skill and by eval Assertion-by-assertion grading evidence with judge reasoning Full target output, side by side for with_skill and without_skill Prompt and judge prompt details Timing and token usage Tool calls when present Use --report-output (or report.output in YAML) to choose where the report lands. agentskills.io compatibility Implements the agentskills.io specification end to end: SKILL.md YAML frontmat