Multi-Agent Patterns MOC
Map of Content for recurring architectural patterns in multi-agent orchestration. This MOC unifies patterns from ADK, OpenAI Swarm, and A2A.
Use this map when the question is "what coordination pattern am I looking at?" rather than "which framework do I use?" The notes here are the reusable shapes that sit underneath specific SDKs and protocols.
🏗️ Structural Patterns
These patterns define the stable topology of a multi-agent system.
- pattern-agent-as-tool - Treating a peer agent as a callable function/tool.
- multi-agent-systems - Tree-based orchestration where leads manage specialists.
- workflow-agents - Deterministic (Sequential, Parallel, Loop) orchestration.
🤝 Collaboration & Handoff
These patterns define how work moves across agent boundaries.
- pattern-dynamic-delegation - LLM-driven runtime decision to delegate to a peer.
- hermes-subagent-delegation - Hermes's concrete in-turn child-agent implementation of isolated delegation and fan-out.
- pattern-progressive-handoff - Gradual transfer of context and control.
- pattern-state-transfer - Methods for passing
context_variablesand state across handoffs.
🗄️ Shared State & Memory
These patterns define how agents share knowledge across boundaries without direct coupling.
- Shared Memory Blackboard - RAG-backed store where agents coordinate by writing and retrieving semantically; the modern form of the classical blackboard architecture.
- hermes-kanban - Durable SQLite-backed task board where named Hermes profiles coordinate through tasks, comments, links, and dispatcher state.
🛡️ Control & Safety
These patterns define how orchestration stays bounded, auditable, and interruptible.
- pattern-capability-gating - Restricting agent access based on validated tokens or lattices.
- pattern-human-in-the-loop - Inserting manual approval steps into automated flows.
- pattern-parallel-fan-out - Scaling request processing across multiple concurrent workers.
- Diversity-Aware Orchestration - Mitigating mode collapse in planning by eliciting a distribution of potential next steps.
- Maker-Checker / Debate - Separating generation and critique into independent agents to mitigate hallucination.
Where To Start
- If your system has a lead agent calling specialists, start with multi-agent-systems and pattern-agent-as-tool.
- If your question is routing or ownership transfer, start with pattern-dynamic-delegation and pattern-progressive-handoff.
- If your concern is trust or approvals, start with pattern-capability-gating and pattern-human-in-the-loop.
- If your concern is concurrency, start with pattern-parallel-fan-out and workflow-agents.