Multi-Agent Orchestration Planner

Multi-agent orchestration is the coordination of multiple AI agents to accomplish complex tasks that are too large, too specialized, or too risky for a single agent. An orchestrator decomposes a goal into sub-tasks, assigns them to specialized agents, manages information flow between agents, handles failures, and synthesizes final results. Examples: a research orchestrator that uses researcher, critic, and writer agents in sequence.

Use multiple agents when: (1) The task is too long for one context window. (2) Different sub-tasks require different tools or expertise. (3) Parallel execution would meaningfully speed up the work. (4) Independent verification is needed (one agent generates, another reviews). (5) Failure isolation is important — one failing agent shouldn't crash everything. Use a single agent for tasks that fit in one context and don't benefit from specialization.

Key patterns: (1) Sequential/pipeline — agent A passes output to agent B, which passes to agent C. Simple but slow. (2) Parallel — independent agents work simultaneously, orchestrator merges results. Fast for independent tasks. (3) Hierarchical — manager agent delegates to worker agents, can recursively delegate. (4) Critic/reviewer — one agent generates, another critiques and improves. (5) Role-based crew — agents have persistent roles (researcher, writer, editor) and collaborate on shared goals.

Failure strategies: (1) Retry with backoff — retry the failed agent 2-3 times before escalating. (2) Fallback agent — a simpler/cheaper backup agent handles the task if the primary fails. (3) Human escalation — route to human review when agents fail after max retries. (4) Graceful degradation — return partial results with a clear indication of what failed. (5) Dead letter queue — save failed tasks for later retry or analysis.

Popular frameworks: LangGraph (graph-based state machines, from LangChain), CrewAI (role-based crews with delegation), AutoGen (Microsoft, conversation-based multi-agent), Swarm (OpenAI, lightweight agent handoffs), LlamaIndex Workflows, Agno (formerly Phidata), and Pydantic AI. Claude Code itself operates as an agent and can be orchestrated via its SDK. For production, LangGraph and CrewAI are most mature.