Multi-Agent Patterns
Multi-Agent Patterns
Designing systems where multiple AI agents collaborate, delegate, and coordinate
Summary
As systems grow complex, single agents give way to multi-agent patterns where specialists collaborate. Five primary patterns exist: sequential (pipeline), concurrent (fan-out), handoff (routing), group chat (collaborative refinement), and dynamic orchestration (emergent planning). Most production systems combine patterns with a deterministic backbone and AI at decision points.
- Sequential pipeline: output of one agent is input to the next
- Concurrent fan-out: same input to multiple agents, aggregate results
- Handoff routing: one agent routes to exactly one specialist
- Group chat: multiple agents share a thread, manager controls order
- Dynamic orchestration: manager decomposes goal, assigns tasks, revises plan
- 2026 consensus: deterministic routing + AI at judgment points
As AI systems grow more capable, single-agent architectures give way to multi-agent patterns where specialized agents collaborate on complex tasks. Your system needs to support this by enabling clean delegation, shared context, and coordination.
What This Section Covers
- Orchestrator-worker patterns — a central agent that decomposes tasks and delegates to specialized sub-agents
- Agent-to-agent delegation — how one agent hands off work to another with context, constraints, and expected outcomes
- Shared context management — passing conversation history, tool results, and accumulated state between agents without losing information
- Fan-out and aggregation — dispatching parallel work to multiple agents and combining their results coherently
- Capability advertising — how agents describe what they can do so orchestrators can route tasks to the right specialist
- Conflict resolution — handling disagreements between agents, competing writes, and inconsistent state
- Observability — tracing multi-agent interactions, measuring latency across agent hops, and debugging coordination failures
See also
- Cookbook: Agentic Loop as Commodity — building individual agents with ToolLoopAgent
- Cookbook: Autonomous Background Agents — agents for proactive tasks, batching, and scheduling