**A multi-agent system** runs two or more separate LLM-driven agents that work together on a task. Each agent has its own model call, its own prompt and persona, and its own set of tools. They coordinate by sending messages to each other, not by sharing one big prompt.
The word that trips people up is multi. It means multiple agents, not multiple tools. A single agent that has access to twenty tools is still one agent: there is one LLM context per turn, one persona, one decision-maker. Add a second separate LLM context with its own role and you now have two agents.
The most common multi-agent setups use role specialization. One agent plans, another researches, another writes, another reviews. Each plays a focused role with a tighter prompt and a more relevant tool set than a single jack of all trades agent could. Coordination happens through message passing managed by the runtime.
Three frameworks dominate in 2026. CrewAI models a crew of role-playing agents (planner, researcher, writer). AutoGen models agents as participants in a multi-party conversation. LangGraph models the coordination as a typed state graph where nodes can be different agents. All three implement the same fundamental idea: multiple separate LLM contexts, coordinated by a runtime, talking through messages.
**A multi-agent system** runs two or more separate LLM-driven agents that coordinate, typically with distinct roles, prompts, and tool sets. They communicate by passing messages, not by sharing one prompt. Each agent has its own model context and decision boundary.
The definitional line is the agent count, not the tool count. A single agent with a rich tool registry is still one agent: one LLM call per turn, one persona, one context window. The minute you spin up a second LLM call with its own role description and its own state, you have a multi-agent system. This is structural, not a matter of perceived intelligence.
Multi-agent designs trade complexity for specialization. Each agent gets a tighter prompt and a more focused tool registry, which often produces better behavior than a single jack of all trades agent. Common patterns include planner-executor (one agent decomposes the task, another runs the steps), researcher-writer (one gathers, one synthesizes), and debate (two agents argue, a third decides). The cost is coordination overhead: more model calls, more orchestration logic, more places for the system to deadlock or loop.
Three production frameworks dominate. **CrewAI** models a crew of role-playing agents with built-in delegation and shared memory. **AutoGen** treats agents as participants in a multi-party conversation, with a group-chat manager driving turn order. **LangGraph** lets each node in a state graph be a different agent, with explicit edges describing coordination patterns. All three implement the same fundamental pattern: separate LLM contexts coordinated by a runtime through message passing.
Quick reference
TL;DR
A multi-agent system runs two or more separate LLM-driven agents that coordinate via message passing, each with its own context, persona, and tool set. The 'multi' is in the agent count, not the tool count.
Think of a tiny office. One person can do many jobs if they have the right tools: a laptop, a phone, a calculator. That is a single agent with many tools. Now imagine the office hires a second person. They have their own desk, their own laptop, their own job description, and they can talk to the first person but they cannot read each other's minds. That is a multi-agent system. The 'multi' is about how many separate brains there are, not how many gadgets each brain has. CrewAI is like a small startup with a planner, a researcher, and a writer. AutoGen is like a conference call between specialists. LangGraph is more like a workflow diagram where each box is its own little office.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.
Single-role tasks where one prompt and one tool set can cover the work. Coordination overhead is wasted; a single agent will be faster, cheaper, and easier to debug.
Latency-sensitive interactive flows. Each agent transition adds at least one model round trip; multi-agent systems struggle with sub-five-second response budgets.
Tasks where you cannot define clear, non-overlapping role boundaries. If two agents end up doing the same kind of work, the system devolves into context fragmentation and coordination overhead with no specialization gain.
Quick scan · 2–4 min
Everything important, quickly.
5 min: define multi-agent in terms of separate LLM contexts and message passing, distinguish from single agent with many tools, list two or three coordination patterns, name three production frameworks (CrewAI, AutoGen, LangGraph), and warn about the main gotchas (context fragmentation, deadlock, cost compounding).
Real-world examples
Real products, models, and research that use this idea.
1CrewAI: a 'crew' of role-played agents (planner, researcher, writer, reviewer) collaborating on a content task with built-in delegation and shared memory.
2Microsoft AutoGen: agents as participants in a multi-party conversation, with a group-chat manager driving turn order and a Codex-style executor agent for tool runs.
3LangGraph: production deployments where each graph node is an agent with its own prompt and tools, and explicit edges describe coordination patterns and recursion bounds.
4Anthropic's research on building effective agents recommends starting single-agent and only adding agents when the role separation has clear value, not by default.
Interviewer follow-ups
What an interviewer would ask next. Try answering before peeking at the approach.
QWhen does multi-agent actually beat a single-agent design with conditional role prompting?
Sign in or go Pro to see the approach for this follow-up.
QHow do agents share what they have learned across a multi-agent run?
Sign in or go Pro to see the approach for this follow-up.
QWhat stops two agents from infinitely deferring to each other?
Sign in or go Pro to see the approach for this follow-up.
Don't say this
Red flags and common mistakes that signal junior thinking. Click to expand.
Most common mistake
Calling a single agent with many tools 'multi-agent'. The structural difference is separate model contexts and personas, not tool count; one LLM context window equals one agent.