An Agent Skill for Claude that designs, builds, and debugs multi-agent systems. Turn Claude into a swarm architect that creates cost-efficient, deterministic-first agent pipelines.
Built by Swarm Labs USA | GitHub
This skill teaches Claude to architect multi-agent swarms:
- Design agent roles using 8 archetypes (Perceiver, Classifier, Analyzer, Synthesizer, Validator, Resolver, Planner, Executor)
- Orchestrate agents with 5 communication patterns (Sequential Chalkboard, Fan-Out/Fan-In, Pipeline with Gates, Adversarial Pair, Iterative Refinement)
- Optimize costs with the deterministic-first cascade — try rules before LLMs, start smallest
- Debug multi-agent failures with chalkboard trace analysis
- Validate pipeline definitions for cycles, schema mismatches, and budget violations
- Simulate pipeline execution to verify data flow before building
Anthropic shipped agent teams in Opus 4.6. 123+ plugins in the marketplace. Zero of them teach how to design, orchestrate, or debug multi-agent systems.
This skill fills that gap with production-tested patterns. The architectures and patterns in this skill are extracted from real systems — see Production Validation below.
# From GitHub
/plugin install github:michaelwinczuk/swarm-orchestrator"Design an agent swarm that monitors social media mentions,
classifies sentiment, and drafts responses for negative ones."
"My pipeline has 12 agents costing $2/run. Help me optimize."
"Agent 3 and Agent 5 produce contradictory outputs. Debug this."
"Validate this pipeline definition for issues."
| Archetype | Role | Typically Deterministic? |
|---|---|---|
| Perceiver | Extract structure from raw data | Always |
| Classifier | Route and categorize | Almost always |
| Analyzer | Deep domain analysis | Sometimes |
| Synthesizer | Combine multiple outputs | Rarely |
| Validator | Quality gates | Always |
| Resolver | Handle conflicts | Sometimes |
| Planner | Decompose goals | Sometimes |
| Executor | Carry out actions | Always |
- Sequential Chalkboard — Agents execute in order, read/write shared state. Default choice.
- Fan-Out / Fan-In — Parallel independent analyses merged by a synthesizer.
- Pipeline with Gates — Sequential with quality checkpoints that can halt/redirect.
- Adversarial Pair — Advocate + Critic reviewed by a Resolver. For high-stakes decisions.
- Iterative Refinement — Generate + validate loop with retry limits.
Rules/Code ($0) → Local Model ($0) → API Model ($0.02) → Frontier ($0.10+)
Always try the cheapest approach first. Most swarm operations don't need LLMs.
Shared state that agents read from and write to:
- Agents can READ any previous agent's output
- Agents can only WRITE to their own section
- Each output is immutable once written
- The chalkboard is the single source of truth and your primary debugging tool
These patterns are extracted from systems built with the Swarm Labs ecosystem. Measured results from deployed swarms:
| Swarm | Agents | Deterministic Rate | Latency | Cost/Run |
|---|---|---|---|---|
| Think Tank (research) | 10 | 69 clusters, 77 KGs | ~2s | ~$0.02 |
| Trading (ETH/BTC/USDC) | 5 | Oracle + rules-first | 46ms | $0.00 |
| SBIR Defense (3 swarms) | 12 each | 95%+ deterministic | 46ms | $0.00 |
| Math Swarm | 6 | 100% (SymPy) | 1.9ms | $0.00 |
| Swarm Claw (orchestrator) | 10 swarms | 4-layer cascade | <1ms routing | $0.00 |
The Math Swarm is the strongest proof of the deterministic-first principle. Full results and code:
| System | Math Accuracy | Speed |
|---|---|---|
| Math Swarm (6-agent chalkboard) | 1,079/1,079 (100%) | 1.9ms |
| Qwen2.5-3B (alone) | 52/94 (55%) | 200ms |
| Qwen2.5-7B (alone) | 72/94 (77%) | 300ms |
| Qwen2.5-32B (alone) | 87/94 (93%) | 2,600ms |
A 3B model + deterministic swarm outperforms a 32B model alone. The architecture matters more than the model size.
swarm-orchestrator/
├── SKILL.md # Core skill (loaded by Claude)
├── README.md # This file
├── LICENSE # Apache 2.0
├── scripts/
│ ├── validate_pipeline.py # Pipeline definition validator
│ └── simulate.py # Chalkboard simulation engine
├── references/
│ ├── archetypes.md # 8 agent archetypes with design rules
│ └── examples.md # 4 complete swarm designs + debugging example
└── evals/
└── evals.json # 5 test cases
| Project | What It Does | Status |
|---|---|---|
| Math Swarm | Zero-hallucination computation. 1,079 tests, 100%, 12 categories. | 1,079 tests passing |
| Knowledge Graph Reasoning | 77+ KGs, adversarial validation, deterministic reasoning. | 77 graphs deployed |
| PRISM | Reliability primitives — VotingMesh, Sentinel, checkpoint/replay. | 95 tests passing |
| Bastion | Safety kernel — consensus, verification, SHA-256 audit trails. | Rust + Tokio |
| Swarm Labs USA | Autonomous AI systems for government. | Active |
Apache 2.0 — See LICENSE for details.
Built by Michael Winczuk at Swarm Labs USA