Definition
A multi-agent system coordinates two or more LLM-driven agents that each hold a role, a set of tools, and sometimes their own memory, working together toward a shared goal. Instead of one model trying to do everything in a single context window, work is split: a researcher agent gathers facts, a writer agent drafts, a reviewer agent critiques, and so on. Agents communicate by passing messages, results, or structured handoffs, and an orchestration layer decides who acts next.
Common topologies
Anthropic's orchestrator-workers pattern uses a central agent to break a task into subtasks dynamically, delegate them to worker agents, and synthesise the results. Other arrangements are hierarchical (a planner over executors), sequential (a pipeline of handoffs), or peer-to-peer (agents negotiating). The right shape depends on whether the subtasks are predictable in advance or must be discovered at runtime.
Trade-offs
Splitting work across specialists can improve quality and make each agent simpler to reason about, but it adds latency and cost because every agent is another set of model calls, and errors can compound as messages pass between them. Anthropic's guidance is to reach for multi-agent designs only when a single agent or a fixed workflow genuinely falls short.
Multi-agent systems are built from the same primitives as a single agent, including tool use and the planner-executor pattern. Running such a system on your own hardware keeps the orchestration, the messages between agents, and the data they touch under your control. D-Central documents these architectures for sovereign self-hosters.
In Simple Terms
A multi-agent system coordinates two or more LLM-driven agents that each hold a role, a set of tools, and sometimes their own memory, working together…