Agent Orchestration

Patterns for building systems where multiple AI agents collaborate, specialize, and hand off work.

Orchestration patterns

Router pattern

A central router agent classifies the request and delegates to specialist agents:

• User query → Router → {Code Agent, Research Agent, Writing Agent}

Pipeline pattern

Agents process sequentially, each adding to or refining the output:

• Research Agent → Analysis Agent → Writing Agent → Review Agent

Supervisor pattern

A supervisor agent coordinates workers, checks quality, and decides when to loop:

• Supervisor → Worker A, Worker B → Supervisor reviews → Done or retry

Swarm pattern

Agents operate independently with shared state, converging on a solution:

• Multiple agents explore different approaches; best result wins

Handoff protocol

• Include full context in the handoff — never assume the next agent remembers
• Pass structured data, not natural language summaries, between agents
• Include the task, constraints, and success criteria in every handoff
• Log every handoff for debugging and audit

State management

• Shared state store (database, Redis) for cross-agent coordination
• Each agent writes its output to a known location
• Version state to handle concurrent modifications