Evaluate Agent Results
Evaluates and ranks agent results for an AgentHub session using metric-based commands, an LLM judge, or a hybrid approach. It produces a ranked list of agents with a highlighted winner and justification.
✨ What it does
- Supports metric-based evaluation with configurable eval command and metric.
- Supports LLM judge mode comparing diffs and result posts.
- Ranks by correctness, simplicity, and quality.
- Hybrid mode runs metric eval first, then LLM judge for close results.
- Updates session state to 'evaluating'.
- Provides clear winner and next-step suggestions.
🎯 When to use it
- When you need to score or compare completed AgentHub agents in a session.
- When you want to pick the best result among multiple agents.
- When you need to break ties between top-performing agents.
- When you want to evaluate agents using a specific metric command.
🚀 How to use
Trigger: /hub:eval
Trigger with the slash command /hub:eval optionally followed by a session ID and flags like --judge. It expects an AgentHub session with completed agents and optionally a configured eval command. Example prompts:
/hub:eval
/hub:eval 20260317-143022 --judge
📄 Output: A ranked list of agents with metrics or verdicts, highlighting the winner.
📦 Add this skill to Claude Code
# 1. Get the skills repo
git clone --depth 1 https://github.com/alirezarezvani/claude-skills /tmp/claude-skills
# 2. Copy this skill into your project (or ~/.claude/skills for all projects)
mkdir -p .claude/skills
cp -r /tmp/claude-skills/engineering/agenthub/skills/eval .claude/skills/evalSkill source: engineering/agenthub/skills/eval/SKILL.md
⚠️ Good to know
Requires an AgentHub session with completed agents and proper git branches for diff comparison.
❓ FAQ
How does metric mode work?
It runs a configured eval command in each agent's worktree and ranks results by the specified metric and direction.
What does LLM judge mode consider?
It compares diffs and result posts, ranking by correctness, simplicity, and quality.
What happens after evaluation?
The session state is updated to 'evaluating' and you are prompted to merge the winner with /hub:merge.
🤖 Overview, features, install steps and FAQ were generated from the project's SKILL.md on Sep 4, 2026. Always check the original source before running commands.