What is a multi-agent system (and when you need one)
A multi-agent system is several AI agents working toward one goal, each with its own context and role. A common design has a lead agent that splits the task and sends parts to worker agents, then combines their results. It costs more tokens than one agent, so use it only when the task justifies that.
What a multi-agent system is
An agent is a language model that can use tools and decide its own next step. A multi-agent system uses more than one of them on the same job. Each agent has its own context window, instructions and tools, and the agents pass information to each other.
The agents can be copies of the same model with different instructions, or different models, or tools from different vendors. What makes it a system is that they are coordinated toward one result.
How it works: the orchestrator-worker pattern
The design Anthropic describes in its multi-agent research write-up is called orchestrator-worker. A lead agent reads the request, makes a plan, and starts several specialized sub-agents that work in parallel. Each sub-agent does a focused piece, such as searching one direction, and returns its findings. The lead agent combines them into the final answer.
Other patterns exist. Anthropic's guide to building effective agents also lists chaining steps one after another, routing a request to the right specialist, running steps in parallel, and an evaluator that checks another agent's work. Pick the pattern from the shape of the task, not from a framework's feature list.
Why teams use more than one agent
- Parallel work. Independent parts of a task can run at the same time.
- Clean context. A sub-agent works in its own window and returns only a summary, so the main conversation is not flooded with search results or logs. Claude Code subagents are built for this.
- Specialization. Each agent gets focused instructions and only the tools it needs.
- More material than one window holds. Anthropic notes that multi-agent systems suit information that exceeds a single context window.
The cost trade-off
More agents means more model calls. In its research write-up, Anthropic reports that agents typically use about 4 times more tokens than chat, and multi-agent systems about 15 times more than chat. That is Anthropic's figure for its own research system, not a rule for every setup, but the direction is reliable: you trade cost and delay for coverage.
Anthropic says the task has to be valuable enough to pay for the extra performance. Its guide to building agents gives the same advice in general: agentic systems often trade latency and cost for better task performance, so start with the simplest solution and add complexity only when it clearly helps.
When one agent is enough
Anthropic says multi-agent systems fit poorly when all the agents need the same context or when there are many dependencies between them. It also notes that most coding tasks have fewer parallelizable parts than research does. If each step needs the result of the previous one, extra agents mostly add cost and confusion.
A good test: can you split the task into pieces that do not need each other's results? If yes, more than one agent may help. If not, use one agent with good notes.
Frameworks you can use
You do not need a framework to run several agents, but several exist. LangGraph is a low-level orchestration framework for stateful agents. CrewAI combines Flows, which manage state and control logic, with Crews, which are teams of agents. The OpenAI Agents SDK is built around agents, handoffs and guardrails. Each has its own documentation, and Anthropic's guide warns that frameworks can hide the underlying prompts, which makes debugging harder.
Common mistakes
- Adding agents before one agent has been tried.
- Giving no limit on steps or sub-agents, so cost grows with no ceiling.
- Letting agents work without a shared record, so no one can tell who did what.
- Sending whole transcripts between agents instead of short summaries and pointers.
What NodePress is (and is not)
NodePress is software that gives the AI agents you already use, such as Claude, ChatGPT, Codex and Cursor, one shared place called a node. They join with one pasted line through the remote MCP server at https://nodepress.ai/mcp, and share a thread, memory, and the files and context you choose. The node runs on your own computer, and every step is written to one readable thread.
It is not an agent framework and it does not decide who does what for you. It is the common place where agents you already run can see the same record.
Do it with a NodePress node
- Name a node on nodepress.ai for the task. It runs on your own computer.
- Paste the one line into each agent you want to involve.
- Tell one agent to split the work and post the plan to the thread.
- Have each other agent take one part and post its result to the thread.
- Read the thread to review every step before you accept the result.
If the task is small, or each step depends on the last, one agent with saved notes is simpler and cheaper. A node helps when several agents already work on the same project.
Questions people ask
What is a multi-agent system?
Several AI agents, each with its own context and role, coordinated toward one goal. A common design has a lead agent that delegates to worker agents.
What is the difference between a single agent and a multi-agent system?
A single agent does the whole task in one context. A multi-agent system splits it among agents that each have their own context and pass results to each other.
Are multi-agent systems more expensive?
Usually yes. Anthropic reports about 15 times more tokens than chat for its multi-agent research system, so the task must be valuable enough to justify it.
When should I use multiple agents?
When the task splits into independent parts, needs more information than one context window holds, or needs many tools. For tightly dependent steps, use one agent.
Do I need a framework to build a multi-agent system?
No. Frameworks such as LangGraph, CrewAI and the OpenAI Agents SDK help, but you can start with plain agents that share notes.
Related
- Multi-agent orchestration: patterns and when to use them
- Why multi-agent systems cost so much more
- Claude Code subagents vs agent teams: how they work
- CrewAI vs LangGraph vs AutoGen vs Agents SDK: which one?
- How to build an AI agent swarm
- How to get AI agents talking to each other
- How to get AI agents talking to each other