Hive is a Claude Code skill that turns any complex task into a coordinated swarm of specialist AI agents. Give it a goal; it designs the team, assigns the right model to each role, and delivers a unified result.
Works for code, research, writing, strategy, analysis, creative projects — anything with more than one moving part.
Most multi-agent setups assign the same model to every agent and hope for the best.
Hive is built on a different principle: the right model for the right cognitive job. A researcher synthesizing 50 sources needs Opus. A formatter applying a template needs Haiku. Treating them identically wastes cost on one and breaks quality on the other.
This is Aristotelian telos-matching — every agent's model is chosen by what the role demands, not by habit.
Hive runs seven phases automatically:
| Phase | What Happens |
|---|---|
| 1. Task Analysis | Decomposes the goal into components, scores complexity and coupling |
| 2. Mode Selection | Picks parallel, hybrid wave, or sequential based on coupling score |
| 3. Agent Design | Assigns role + model + harness (Claude vs Codex CLI) to each agent |
| 4. Contract Generation | Opus writes a shared contract — the single source of truth for all agents |
| 5. Execution | Agents run with full contract context; Hybrid Wave agents can message each other |
| 6. Validation | Critic agent checks completeness, coherence, and contract compliance |
| 7. Delivery | Outputs merged in dependency order, targeted fixes applied, final package delivered |
Chosen automatically based on how tightly the components depend on each other:
Coupling < 0.4 → Contract-First Parallel
All agents run simultaneously. No A2A needed.
Speedup: 4–8x vs sequential.
Coupling 0.4–0.7 → Hybrid Wave
Agents grouped in waves. Within a wave: parallel.
Between waves: sequential (frozen outputs passed forward).
A2A messaging enabled for real-time peer coordination.
Speedup: 2–5x vs sequential.
Coupling > 0.7 → Progressive-Layer Sequential
Each layer builds directly on the previous layer's frozen output.
Used when design cannot be fully specified upfront.
In Hybrid Wave mode, agents can message each other directly — without routing through the orchestrator.
Three permitted triggers:
- Interface deviation — output format must change from what the contract specified
- Partial output ready — a section is final and a blocked peer can start consuming it
- Peer-resolvable blocker — a specific gap that only a peer agent can fill
Everything else is resolved through the contract. A2A is the exception channel, not the default one.
The orchestrator (Claude Code) receives all messages automatically.
| Cognitive Demand | Model |
|---|---|
| Deep synthesis, strategy, high-stakes decisions | claude-opus-4.6 |
| Production writing, analysis, implementation | claude-sonnet-4.6 |
| Bulk generation, formatting, pattern application | claude-haiku-4.5 |
| Structured JSON / schema-heavy output | openai/gpt-4.1 |
| Long documents (200k+ tokens) | google/gemini-3.1-pro-preview |
| Real-time / social context | x-ai/grok-3 |
Non-Anthropic models run via Codex CLI with an AI Gateway passthrough.
hive/
├── SKILL.md # Main orchestration logic (7 phases)
├── resources/
│ ├── task-taxonomy.md # 5 task families, complexity scoring, coupling formula
│ ├── agent-roster.md # Universal role catalog with telos descriptions
│ ├── model-harness-guide.md # Model selection rules + harness assignment
│ └── coordination-modes.md # Parallel / Hybrid Wave / Sequential specs + A2A rules
└── templates/
├── agent-brief.md # Per-agent prompt template with A2A section
└── crucible-contract.md # Shared contract template (Section 8: A2A protocol)
Resources are loaded on demand — only the files relevant to the current phase are read.
Use Hive when the task has 3 or more components that must interoperate or when it's too large for one agent's context window.
Good fits:
- "Research the EV market and write a 20-page strategy report"
- "Build a full-stack app with tests and documentation"
- "Create a pitch deck + financial model + executive summary"
- "Analyze this dataset and produce visualizations + narrative"
Not a good fit:
- Single-step tasks (just use Agent directly)
- Quick fixes or one-file changes
- Exploratory research with no concrete deliverable
Copy to your Claude Code skills directory:
# Via ClawHub
/install hive
# Manual
cp -r hive/ ~/.claude/skills/hive/Then trigger with: /hive, multi-agent, agent swarm, or any prompt that asks for a comprehensive multi-part output.
| Anti-Pattern | Fix |
|---|---|
| Same model for all agents | Telos-match every role individually |
| Partial contract per agent | Every agent receives the full contract |
| Sequential execution when coupling < 0.4 | Use parallel — 4–8x speed gain |
| A2A messaging in Parallel mode | A2A is Hybrid Wave only |
| Anonymous agents in Hybrid Wave | Use TeamCreate + name= for A2A to work |
| Messaging peers for status updates | Only message for the 3 permitted triggers |
"The excellence of a thing is the activity by which it fulfills its function well." — Aristotle, Nicomachean Ethics I.7
A model's excellence is not its benchmark score — it's how well it performs the specific task it was designed for. The virtuous orchestrator doesn't use the most powerful model; they use the most appropriate one. Hive's entire architecture is built on this principle.
Built by HappyCapy Research