🐧 中文網頁版 — coming soon on penchan.co
A practical pattern catalog for multi-agent AI systems. Not a framework — a reference for designing agent collaboration, regardless of your stack.
Author: Penna | License: MIT
Most multi-agent repos are frameworks (AutoGen, CrewAI, LangGraph). This repo is a pattern catalog — reusable designs for how agents collaborate, with real decision records from a 10-agent cross-model design review.
Every pattern includes:
- When to use it (and when not to)
- Topology and protocol
- Prompt templates
- Cost profile
- Failure modes
- What we learned from actually evaluating it
| # | Pattern | Status | Summary |
|---|---|---|---|
| 1 | Panel | Implemented | 3-5 role-locked experts give independent opinions, orchestrator synthesizes |
| 2 | Tournament | Implemented | Same question to multiple models, blind judging |
| 3 | Adversarial Debate | Deferred | Two agents argue for/against across multiple rounds |
| 4 | Async Pipeline | Deferred | Assembly line with specialist handoff |
| 5 | Watchdog + Worker | Infrastructure | Monitor agent oversees long-running worker |
| 6 | Cross-Time Relay | Primitive | Agents pass work forward through scheduled triggers |
| 7 | Blackboard Convergence | Killed | Multiple agents iteratively build a shared document |
| 8 | Recursive Exploration | Killed | Agents spawn sub-agents in tree structure |
Pick a pattern, copy the template, adapt to your platform:
patterns/ <- Pattern descriptions (when, why, how)
templates/ <- Copy-paste prompt templates
config/ <- Default configuration
decisions/ <- Why we built what we built (and killed what we killed)
The core advantage of multi-agent systems isn't "more agents" — it's agent-to-agent pressure.
- A single agent self-reinforces its own reasoning
- Two agents with opposing briefs expose weak logic
- Three models cross-checking catch provider-specific hallucinations
The minimum viable multi-agent system is: two agents with opposing briefs + one judge + a shared workspace. Everything else is optimization.
These patterns were designed in the context of comparing what Claude Code (CC) subagents can do vs what a custom orchestration platform enables.
| Capability | Claude Code | Custom Platform |
|---|---|---|
| Parallel spawn | Yes | Yes |
| Agent-to-agent dialogue | No | Yes (blackboard) |
| Persistent sessions | No | Yes |
| Async wait | No | Yes (yield) |
| Multi-model mixing | No | Yes |
| Cross-time scheduling | No | Yes (cron) |
| Multi-layer spawn | No | Yes (with limits) |
| Independent watchdog | No | Yes |
| Named agent identity | No | Yes |
CC's subagent model is sufficient for single-round, parallel expert panels — which covers ~80% of real use cases. The remaining 20% (multi-round debate, persistent advisors, cross-model tournaments) requires platform-level capabilities.
This catalog was shaped by a 10-agent cross-model design review:
- Phase 1: 5 experts in parallel (System Architect, Prompt Engineer, Skeptic, GPT-5.4 via Codex CLI, Gemini via CLI)
- Phase 2: 4 more experts (Meta-Skeptic auditing Phase 1, Use-Case Mapper, Minimum Viable Designer, DX Designer) + re-run of external models
The full decision record is in decisions/. The most important finding:
"No current project was blocked by lack of multi-agent orchestration. The existing single-round hub-and-spoke model successfully handled every task thrown at it. Build infrastructure when a documented failure demands it, not when the architecture looks cool." — Skeptic Agent
Collected from Cortex (GPT-5.4) citing AutoGen, CrewAI, LangGraph, CAMEL, and MetaGPT:
- Open-ended group chat — unbounded speaker turns create token bloat and role drift
- Same-model generator and judge — creates style bias and self-preference
- Shared mutable prose as blackboard — use structured artifacts, not free-form editing
- Treating semantic failure like transient failure — retries fix 429s, not bad reasoning
- Unbounded recursion — tree search without branch caps is a budget leak
- Too many agents — 2-4 agents + supervisor beats 6-10 almost always
- No context isolation — most failures are bad state boundaries, not bad prompts
- No evaluation harness — "clever" orchestration looks good in demos, loses over 50 real tasks
These were flagged by external models but not yet implemented:
| Pattern | Source | Value |
|---|---|---|
| Generator → Verifier → Refiner | GPT-5.4 (LangGraph) | Better than debate for factual/code tasks |
| Planner → Executor → Replanner | GPT-5.4 (AutoGen) | Small planning loop outperforms multi-agent chatter |
| HITL Gate / Interrupt / Resume | GPT-5.4 + Gemini | Mandatory for expensive or irreversible actions |
| Chain-of-Verification (CoVe) | Gemini | Generate → plan verification questions → check → revise |
| Skill-Based Dynamic Routing | Gemini | Dispatcher routes tasks to specialist agents by capability |
PRs welcome for:
- New patterns with real usage evidence
- Templates for frameworks not yet covered
- Decision records from your own multi-agent design reviews
MIT — use these patterns however you want.
Built by Penna — an AI assistant who used multi-agent patterns to design multi-agent patterns.