AgentHub — Multi-Agent Collaboration — Development

AgentHub is a multi-agent collaboration skill that spawns N parallel subagents to compete on the same task, each working in an isolated git worktree. The coordinator evaluates their results using metrics or an LLM judge, then merges the winning branch. It produces a merged solution with archived losing attempts, all tracked in a git repository.

engineering claude-code

📂 View skill on GitHub ↗ 🚀 How to use

✨ What it does

Spawns N parallel subagents in isolated git worktrees.
Supports agent templates for optimizer, refactorer, test-writer, and bug-fixer patterns.
Provides a message board for coordinator-agent communication.
Evaluates results via metric-based ranking, LLM judge, or hybrid mode.
Merges the winning branch and archives losers with git tags.
Tracks session state and DAG with scripts for analysis and cleanup.

🎯 When to use it

When you want multiple approaches tried in parallel for code optimization, content variation, or research exploration.
When you need to compare different solutions or strategies and select the best one.
When you want to run a tournament or fan-out of agents on a single task.
When you need to A/B test copy or generate content variations.
When you want to explore multiple strategies and have them compete.

🚀 How to use

Trigger: /hub:spawn

Trigger AgentHub with slash commands like /hub:hub-init, /hub:spawn, /hub:eval, /hub:merge, or use the one-shot /hub:run. It activates on phrases like 'try multiple approaches' or 'have agents compete'. Inputs include a task description, number of agents, and evaluation criteria. Example prompts:

/hub:run task="Optimize the sorting algorithm" agents=3 eval="pytest bench.py --json" metric=p50_ms direction=lower

Try multiple approaches to reduce API latency and have agents compete.

📄 Output: A merged winning branch with archived losing attempts, plus session logs and board posts.

📦 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/agenthub .claude/skills/agenthub

Skill source: engineering/agenthub/skills/agenthub/SKILL.md

⚠️ Good to know

Requires a git repository and assumes the coordinator is a Claude Code session with the Agent tool available.

❓ FAQ

How do agents communicate with each other?

Agents do not see each other's work and do not communicate with each other; they only write to the message board for the coordinator to read.

What happens if no agent improves over the baseline?

The session is archived with no winner, and the coordinator may suggest different approaches.

Can I use both metric and LLM judge evaluation?

Yes, hybrid mode runs metric first and uses LLM judge to break ties if top agents are within 10% of each other.

🤖 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.