Agenthub

AgentHub — Multi-Agent Collaboration

Spawn N parallel AI agents that compete on the same task. Each agent works in an isolated git worktree. The coordinator evaluates results and merges the winner.

Slash Commands

Agent Templates

When spawning with --template, agents follow a predefined iteration pattern:

Templates are defined in references/agent-templates.md.

When This Skill Activates

Trigger phrases:

• "try multiple approaches"
• "have agents compete"
• "parallel optimization"
• "spawn N agents"
• "compare different solutions"
• "fan-out" or "tournament"
• "generate content variations"
• "compare different drafts"
• "A/B test copy"
• "explore multiple strategies"

Coordinator Protocol

The main Claude Code session is the coordinator. It follows this lifecycle:

INIT → DISPATCH → MONITOR → EVALUATE → MERGE

1. Init

Run /hub:hub-init to create a session. This generates:

• .agenthub/sessions/{session-id}/config.yaml — task config
• .agenthub/sessions/{session-id}/state.json — state machine
• .agenthub/board/ — message board channels

2. Dispatch

Run /hub:spawn to launch agents. For each agent 1..N:

• Post task assignment to .agenthub/board/dispatch/
• Spawn via Agent tool with isolation: "worktree"
• All agents launched in a single message (parallel)

3. Monitor

Run /hub:hub-status to check progress:

• dag_analyzer.py --status --session {id} shows branch state
• Board progress/ channel has agent updates

4. Evaluate

Run /hub:eval to rank results:

• Metric mode: run eval command in each worktree, parse numeric result
• Judge mode: read diffs, coordinator ranks by quality
• Hybrid: metric first, LLM-judge for ties

5. Merge

Run /hub:merge to finalize:

• git merge --no-ff winner into base branch
• Tag losers: git tag hub/archive/{session}/agent-{i}
• Clean up worktrees
• Post merge summary to board

Agent Protocol

Each subagent receives this prompt pattern:

You are agent-{i} in hub session {session-id}.
Your task: {task description}

Instructions:
1. Read your assignment at .agenthub/board/dispatch/{seq}-agent-{i}.md
2. Work in your worktree — make changes, run tests, iterate
3. Commit all changes with descriptive messages
4. Write your result summary to .agenthub/board/results/agent-{i}-result.md
5. Exit when done

Agents do NOT see each other's work. They do NOT communicate with each other. They only write to the board for the coordinator to read.

DAG Model

Branch Naming

hub/{session-id}/agent-{N}/attempt-{M}

• Session ID: timestamp-based (YYYYMMDD-HHMMSS)
• Agent N: sequential (1 to agent-count)
• Attempt M: increments on retry (usually 1)

Frontier Detection

Frontier = branch tips with no child branches. Equivalent to AgentHub's "leaves" query.

python scripts/dag_analyzer.py --frontier --session {id}

Immutability

The DAG is append-only:

• Never rebase or force-push agent branches
• Never delete commits (only branch refs after archival)
• Every approach preserved via git tags

Message Board

Location: .agenthub/board/

Channels

Post Format

---
author: agent-1
timestamp: 2026-03-17T14:30:22Z
channel: results
parent: null
---

## Result Summary

- **Approach**: Replaced O(n²) sort with hash map
- **Files changed**: 3
- **Metric**: 142ms (baseline: 180ms, delta: -38ms)
- **Confidence**: High — all tests pass

Board Rules

• Append-only: never edit or delete posts
• Unique filenames: {seq:03d}-{author}-{timestamp}.md
• YAML frontmatter required on all posts

Evaluation Modes

Metric-Based

Best for: benchmarks, test pass rates, file sizes, response times.

python scripts/result_ranker.py --session {id} \
  --eval-cmd "pytest bench.py --json" \
  --metric p50_ms --direction lower

The ranker runs the eval command in each agent's worktree directory and parses the metric from stdout.

LLM Judge

Best for: code quality, readability, architecture decisions.

The coordinator reads each agent's diff (git diff base...agent-branch) and ranks by:

1. Correctness (does it solve the task?)
2. Simplicity (fewer lines changed preferred)
3. Quality (clean execution, good structure)

Hybrid

Run metric first. If top agents are within 10% of each other, use LLM judge to break ties.

Session Lifecycle

init → running → evaluating → merged
                            → archived (if no winner)

State transitions managed by session_manager.py:

Proactive Triggers

The coordinator should act when:

Installation

# Copy to your Claude Code skills directory
cp -r engineering/agenthub ~/.claude/skills/agenthub

# Or install via ClawHub
clawhub install agenthub

Scripts

Related Skills

• autoresearch-agent — Single-agent optimization loop (use AgentHub when you want N agents competing)
• self-improving-agent — Self-modifying agent (use AgentHub when you want external competition)
• git-worktree-manager — Git worktree utilities (AgentHub uses worktrees internally)