agenthub

---
name: agenthub
description: Use when multiple approaches should compete in parallel — code optimization, content variants, or research exploration — then be scored and the best one adopted.
category: Engineering
version: 1.0.0
tools: []
---

# AgentHub — Multi-Agent Collaboration

Coordinate multiple teammates working the same task independently and in parallel, then compare their results and adopt the strongest one.

## Lifecycle

INIT → DISPATCH → MONITOR → EVALUATE → MERGE

## Coordinator Protocol

1. **Init** — Define the task, how many teammates will attempt it, and the evaluation criteria (metric or judge).
2. **Dispatch** — @mention each teammate with the identical task assignment, launched together so they run in parallel.
3. **Monitor** — Track progress as each teammate reports back.
4. **Evaluate** — Score each result using metric mode, judge mode, or hybrid (see below).
5. **Merge** — Adopt the winning result, summarize why it won, and keep the rest for reference.

## Agent Templates

| Template | Pattern | Use Case |
|----------|---------|----------|
| Optimizer | Edit → evaluate → keep/discard → repeat | Performance, latency, size |
| Refactorer | Restructure → test → iterate until passing | Code quality, tech debt |
| Test-writer | Write tests → measure coverage → repeat | Test coverage gaps |
| Bug-fixer | Reproduce → diagnose → fix → verify | Bug fix approaches |

## Agent Protocol

Each teammate receives the same task description and works independently:
1. Understand the assignment.
2. Do the work — research, draft, or implement — iterating as needed.
3. Report a result summary: the approach taken, what changed, the measured metric (if any), and a confidence level.

Teammates never see each other's work in progress and never coordinate with one another directly — only the coordinator sees all results together.

## Evaluation Modes

**Metric-based** — best for benchmarks, test pass rates, response times, or any quantifiable output. Compare each result's metric value directly against the others and against a baseline. Example: agent-2 scored 142ms versus a 180ms baseline, a 38ms improvement — the clear winner on this criterion.

**LLM judge** — best for code quality, readability, writing quality, or architecture decisions. Rank results by: (1) correctness — does it solve the task; (2) simplicity — less unnecessary change is better when correctness ties; (3) quality — clean execution and good structure.

**Hybrid** — run the metric comparison first. If the top results are within 10% of each other, break the tie with an LLM judge pass.

## DAG Model

Treat every attempt as a node in a directed acyclic graph: each teammate's attempt branches off the shared starting point, and a retry creates a new attempt rather than overwriting the old one. Nothing is discarded until a winner is chosen — every approach stays inspectable so a losing idea can be revisited later. The current frontier is simply the set of attempts with no follow-up work yet.

## Proactive Triggers

| Signal | Action |
|--------|--------|
| All teammates failed | Summarize the failures and suggest retrying with different constraints |
| No improvement over baseline | Note it and suggest different approaches |
| A round is stalled | Check in with the teammate and consider timing it out |