Skip to content

Latest commit

 

History

History
43 lines (29 loc) · 1.99 KB

File metadata and controls

43 lines (29 loc) · 1.99 KB

Hwamony Agent System Evaluator

Turn a running agent system into rubric-backed evidence, stored evaluation artifacts, and the next improvement backlog.

This skill treats AI agent evaluation as a repeatable workflow, not a one-off opinion. It reads implementation artifacts, writes or updates the test plan, initializes run folders, logs inputs and outputs, scores results with a rubric, and captures the next improvement loop.

🚀 Start Here

  • Use $hwamony-agent-system-evaluator to create a test plan for my research agent with run folders, rubric criteria, and failure notes.
  • Use $hwamony-agent-system-evaluator to score 5 sample runs, preserve the evidence, and summarize the biggest weaknesses for the next implementation pass.
  • Use $hwamony-agent-system-evaluator to turn these runs into a backlog of prompt, routing, and implementation fixes.

✨ Why It Is Different

  • it focuses on per-run evidence, not vague quality judgments
  • it stores evaluation artifacts so iteration has a paper trail
  • it combines scoring, notes, and user feedback into one loop
  • it is built to feed implementation backlogs, not just generate a report

🧭 Use It When

  • your agent or prompt system already exists and you want to test it properly
  • you need run-by-run evidence instead of vague impressions
  • you want rubric-based scoring and structured feedback
  • you are preparing for the next implementation iteration

📦 What You Get

  • per-run artifacts under 07-runs/run-###/
  • evaluation summaries
  • rubric-based scores
  • feedback logs
  • a clearer handoff back to implementation

🧠 Best Fit

  • systems that already run but are still improving
  • teams that need evidence before changing prompts, code, or routing
  • workflows where evaluation needs to create the next implementation backlog

⚠️ Boundary

This skill is not for giving a thumbs-up or thumbs-down impression. If the system does not exist yet, use the builder or implementer first and come back when there is something real to evaluate.