Multi-Agent Systems: Design, Orchestration & Decomposition Patterns

A multi-agent system is a set of autonomous agents that interact, each with its own role, context and tools, to solve problems that are hard for one agent. In LLM systems the agents are model-driven and coordinate through messages, shared state or a controller.

Common patterns are sequential pipelines, concurrent or parallel fan-out, a supervisor or orchestrator that routes work, hierarchical teams, handoffs between agents, group chat or debate, and decentralised swarms. Microsoft, Google, AWS and Anthropic each describe overlapping sets.

Task decomposition breaks a goal into subtasks that agents can handle. Patterns include chain-of-thought, least-to-most and plan-and-solve prompting, ReAct loops, orchestrator-workers, map-reduce style parallel work and adaptive decomposition that splits a task only when an agent fails.

Use several agents when work splits into independent parts, needs different tools or permissions, or exceeds one context window. A single agent with good tools is often simpler, cheaper and more reliable, so add agents only when a measured need appears.

In the supervisor pattern a central agent decides which specialist runs next and combines results. In the handoff pattern one agent passes control, and the relevant context, to another, with no permanent central controller.

Research on failures groups them into poor task specification, misalignment between agents and weak verification. Extra agents also add token cost, latency and conflicting decisions, so tracing, evaluation and guardrails matter.

Agents pass messages in a shared conversation, hand off through tool calls, or share state such as a blackboard or memory store. Open protocols such as MCP for tools and A2A for agent-to-agent calls are standardising this.

The section 'Related Protocol References' lists A2A Protocol, Model Context Protocol, AGNTCY.

Links were checked in November 2026. Some sites block automated checks and may need a manual click. This page is a resource list, not engineering advice, and vendor guides reflect their own products.