Add swarm-orchestrator skill

[michaelwinczuk]

Summary

Adds a swarm-orchestrator skill that teaches Claude to design, build, and debug multi-agent systems.

• 8 agent archetypes — Perceiver, Classifier, Analyzer, Synthesizer, Validator, Resolver, Planner, Executor — with design rules and deterministic-first guidance
• 5 communication patterns — Sequential Chalkboard, Fan-Out/Fan-In, Pipeline with Gates, Adversarial Pair, Iterative Refinement
• Deterministic-first cascade — Try rules → local model → API → frontier. Typical 50-82% deterministic solve rate
• Chalkboard protocol — Shared immutable state for agent communication and debugging
• Cost optimization — Per-agent and per-pipeline budgets, deterministic ratio tracking
• Debugging methodology — Chalkboard trace analysis, common failure patterns, health metrics

Included Files

• SKILL.md — Core skill instructions (~300 lines)
• scripts/validate_pipeline.py — Pipeline definition validator (cycles, schemas, budgets)
• scripts/simulate.py — Chalkboard simulation engine for dry runs
• references/archetypes.md — Detailed archetype guide with decision matrix
• references/examples.md — 4 complete swarm designs + debugging walkthrough
• evals/evals.json — 5 test cases

Why This Skill

Anthropic shipped agent teams in Opus 4.6. 123+ plugins in the marketplace. Zero teach multi-agent orchestration. This fills the biggest architectural gap — how to design agent systems that are debuggable, cost-efficient, and deterministic-first.

Companion to knowledge-graph-reasoning (#888) — structured knowledge + agent orchestration = complete autonomous systems.

Test plan

• Verify SKILL.md frontmatter passes validation
• Test swarm design from natural language requirement (eval Reorganize the example skills #1)
• Test debugging contradictory agent outputs (eval Add initial Agent Skills Spec #2)
• Test cost optimization recommendations (eval Adding details about document skills to README #3)
• Test domain-specific swarm design (eval Add 3rd Party notices #4)
• Test pipeline validation with intentional errors (eval Add Claude Code Marketplace plugins #5)

Built by Swarm Labs USA | Standalone repo

🤖 Generated with Claude Code