Opens in a new tab
needsahuman.

Multi-agent systems: when AI agents work as a team

Multi-agent systems split one job between several AI agents that work like a small team. This lesson explains how they work, where you will meet them, and when they are worth the extra cost. In this lesson What a team of AI agents is: a lead that plans, specialists that work in parallel, and a…

Multi-agent systems split one job between several AI agents that work like a small team. This lesson explains how they work, where you will meet them, and when they are worth the extra cost.

In this lesson

  • What a team of AI agents is: a lead that plans, specialists that work in parallel, and a reviewer that checks.
  • The two main patterns: several agents trying the same task, and a lead splitting a job into parts.
  • Where you will meet AI teams in products from the main AI makers, as they describe them.
  • What it costs, and when it is worth it.
  • Your role: the manager who sets the goal, the limits and the sign-off.

The team idea, and its limits

Lesson 7 described one agent working through a task step by step. Some systems split a job between several agents instead, a bit like a small team: a lead agent (sometimes called an orchestrator) plans the work and hands out parts, specialist agents do those parts, often at the same time, and a reviewer checks the result before it comes back to you.

The org chart below is drawn in the same way as the team charts in Lesson 5, so you can see the same shape used for people and for AI. People are amber; AI agents are ink and grey.

Org chart: you, a person, at the top, setting the goal, the limits and the sign-off. Under you, a lead agent that plans, splits and merges the work. Under the lead, four specialist agents in a row: research, drafting, data and checking. All their work comes together at a sign-off by a person before anything leaves the business.

  • You: set the goal, the limits, the sign-off
    • Lead agent: plans, splits, merges
      • Research agent
      • Drafting agent
      • Data agent
      • Checking agent
Sign-off before anything leaves
The same chart shape as a team of people, with you as the manager. Agents do not remember yesterday unless built to, and do not push back, so the goal, the limits and the final sign-off stay with a person.

The metaphor helps, but do not take it too far. AI agents are not colleagues:

  • They do not remember yesterday unless the system is built to keep notes.
  • They do not know when to push back. A person might say “this brief doesn’t make sense”; an agent will usually do its best with it.
  • They do not own the result. The business does, which means you do (Lesson 10).

That is why the chart has you at the top, setting the goal and the limits, and a sign-off box at the bottom before anything leaves the business.

Two patterns in multi-agent systems

Two patterns. Parallel: one question goes to several agents that each try the same task; a lead agent compares and merges their answers into one answer, which you check. Division of labor: a lead agent splits the job; research, drafting and data agents each do a part; a reviewer agent checks; the parts are merged into one result; a person signs off.

Parallel: several agents try the same task

One question
Agent 1 tries it
Agent 2 tries it
Agent 3 tries it
Compare and mergelead agent
Answeryou check

Division of labor: each agent does a part

Lead agent splits the job
Research agent
Drafting agent
Data agent
Reviewer agent checks
Merged result
A person signs off
Both patterns run several agents, so both cost several times one chat. Worth it for broad research and big multi-part jobs; not for simple tasks, where it adds cost and new ways to go wrong.

Parallel: several agents try the same task

Several agents work on the same question at once, independently. Their answers are compared, and the best one is chosen or the answers are merged. The idea is the same as asking three people for an estimate and comparing: where they agree you can be more confident, and where they differ you know to look harder.

Division of labour: a lead splits the job

A lead agent breaks a big task into parts and sends each part to a specialist agent. For example, to prepare a report on a new market, one agent might research competitors, another regulations, another prices, while a fourth checks the sources. The lead then combines their work.

Anthropic, which makes Claude, has described how its own research feature works this way: “The lead agent analyzes it, develops a strategy, and spawns subagents to explore different aspects simultaneously” (Anthropic). In its internal tests, “a multi-agent system with Claude Opus 4 as the lead agent and Claude Sonnet 4 subagents outperformed single-agent Claude Opus 4 by 90.2%” on research tasks. That is the company’s own test of its own product, so treat it as a description of what the design is for rather than independent proof.

Where you will meet AI teams

Most businesses will meet AI teams inside products rather than build them. The main AI makers describe features like these on their own pages (each read on the date shown in the table):

  • Anthropic (Claude). Its Research feature “uses multiple Claude agents to explore complex topics more effectively” (Anthropic). In Claude Code, its coding agent, Claude can hand side tasks to sub-agents: “Each subagent runs in its own context window with a custom system prompt” (Claude Code docs).
  • Google (Gemini). The Gemini Enterprise app offers “Google-made agents like Gemini Notebook and Deep Research,” and lets businesses “build custom agents for your specific workflows in Workflow Builder” (Google Cloud).
  • Microsoft (Copilot). Copilot Studio, for building company agents without code, offers “multi-agent orchestration for complex processes, routing tasks to the right agent” (Microsoft).
  • Mistral AI. Its developer platform lets agents hand work to each other: “Handoff capability to use different agents as part of a workflow” (Mistral docs).
  • OpenAI (ChatGPT). ChatGPT Work takes a goal and “gathers context, makes a plan, and checks in when judgment or approval matters” (OpenAI). For developers, its agent tools let agents “plan and complete tasks using tools, work with other agents” (OpenAI developer docs).
  • xAI (Grok). In the Grok app, “Multiple agents work in parallel for deeper answers on the hardest questions” (xAI). When it launched Grok 4 Heavy in July 2025, xAI described “parallel test-time compute, which allows Grok to consider multiple hypotheses at once” (xAI). Its developer platform has a research mode in which several agents discuss a question and “a designated leader agent is responsible for synthesizing the discussion” (xAI docs).

The two patterns above map onto these: Grok’s parallel agents are the first; Anthropic’s Research, Copilot Studio’s orchestration and the developer tools that hand work between agents are the second.

The full table of agent features, with the date each was read:

AI makers’ agent features, as each describes them on its own pages, with the date we read them. Alphabetical; not a ranking.
MakerAssistantAgent features, as the maker describes themDate read
Alibaba (Qwen)Qwen ChatNot stated on the pages we read.October 6, 2026
AnthropicClaude
  • Research (multi-agent research system): A lead agent plans the research, then starts several sub-agents that search in parallel and report back; a separate agent adds citations. Anthropic says this uses far more tokens than chat: about 4 times for a single agent and about 15 times for a multi-agent system.
  • Research: token use versus chat: Anthropic's figure for the cost of agents and multi-agent systems compared with chat (engineering post published Jun 13, 2025).
  • Research (in the Claude apps): Claude runs several searches that build on each other across the web and connected work apps, and returns an answer with citations. The pricing page lists Research on Pro, Max and paid business plans, not on Free.
  • Claude Cowork (now merging into Claude): Hand Claude a goal and it works across your files and tools, splits big projects into parts that run at the same time, can run on a schedule while your laptop is closed, and asks before taking an action by default. Anthropic says Cowork and chat are merging into one Claude, rolling out to Pro and Max first; the Team plan lists 'Claude Code and Claude Cowork'.
  • Claude in Chrome: A browser extension that reads the page you are signed in to and clicks, types and fills forms, while you approve what happens.
  • Claude Code subagents: In Claude Code, the coding agent, Claude can hand a side task to a specialised subagent that works in its own context with its own tools and permissions and returns a summary; the docs also describe agent teams, a coordinated team of sessions that Claude starts and supervises.
October 6, 2026
DeepSeekDeepSeek (chat at chat.deepseek.com)DeepSeek Harness: An agent harness for developers, in developer preview; the API docs also list guides for running DeepSeek models inside other makers' coding agents.October 6, 2026
GoogleGemini (Gemini app; Gemini Enterprise app for business)October 6, 2026
MetaMeta AI
  • Meta Business Agent: An AI agent that answers a business's customers in chat in their language and tone, lets the business decide when a person steps in, and can give the owner a morning briefing on overnight chats.
  • Meta Business Agent Platform: A platform for larger businesses to build, customise and deploy their Business Agent and connect it to systems such as Shopify and Zendesk so it can take actions.
  • Muse Code: A coding tool for the terminal that uses several agents, built for Muse Spark.
  • Muse Glimmer (open model for local agents): An open model meant for always-on agents that run on your own computer, tuned for tool use, long tasks and recovering from failures.
October 6, 2026
MicrosoftMicrosoft Copilot (formerly Microsoft 365 Copilot); Microsoft Copilot Chat (formerly Microsoft 365 Copilot Chat)
  • Prebuilt Microsoft agents: Researcher and Analyst for research and data analysis with advanced reasoning, and Facilitator for meetings; listed on Microsoft 365 Copilot Business and enterprise plans, with an Agent Store for Microsoft and third-party agents.
  • Agent Builder and Copilot Studio: Agent Builder inside Microsoft Copilot makes quick, simple agents; Copilot Studio builds custom agents, workflows and apps without code, including agents that run tasks or business processes on their own and escalate when needed.
  • Copilot Studio multi-agent orchestration: Connects agents, from Microsoft or partners, so a complex process is routed to the agent with the right knowledge or actions.
October 6, 2026
Mistral AIMistral Vibe (formerly Le Chat)
  • Vibe (assistant and agent): Mistral's assistant now presents itself as an agent for long tasks that uses your knowledge and tools; plans add scheduled prompts, agents or workflows, connectors and custom MCP connectors.
  • Remote coding agents (Vibe for code): Coding agents that run in isolated sandboxes so several tasks run in parallel, with a terminal agent (Vibe CLI) and IDE plugins.
  • Agents and Conversations API: For developers: agents that plan, use built-in tools (code execution, web search, image generation, document library) and hand off to other agents in a workflow.
October 6, 2026
OpenAIChatGPT
  • ChatGPT Work: Give ChatGPT a goal and it gathers context from your apps, files, tools and browser, makes a plan, does the work and checks in for approval; it can also run scheduled and team tasks. Available on all plans on desktop, and on Plus, Pro, Business, Enterprise and Edu on web and mobile.
  • Dots: Always-on agents that keep working between conversations; listed on Pro and on ChatGPT Business Premium seats.
  • Agents SDK and Agents API (developers): Developer tools for building agents that use tools, work with other agents and keep context across steps; the Agents API runs the Codex harness with multi-agent orchestration and MCP support.
October 6, 2026
xAIGrok
  • Multi-agent mode (Grok app): Several agents work on parts of a hard question at the same time, each shows its reasoning, and the results are merged into one cited answer. The page ties multi-agent reasoning to SuperGrok.
  • Grok 4 Heavy / SuperGrok Heavy: xAI described Grok 4 Heavy (July 2025) as using parallel test-time compute to weigh several hypotheses at once, sold through a SuperGrok Heavy tier. The current pricing page lists SuperGrok Heavy as a plan but its text does not describe Heavy as multi-agent.
  • Realtime Multi-agent Research (API, beta): For developers: one request launches 4 or 16 agents that discuss the query, and a leader agent combines their findings into the answer; more agents means deeper research at higher token use and latency.
  • Grok Bot: Always-on agents with their own cloud computer that sign into your tools, finish jobs end to end and come back when something needs your approval; in beta for SuperGrok plans and some Cursor plans, and for enterprises.
  • Workflows in Grok Build: Grok Build, the coding agent, can write and run scripts that split a task across many parallel agents, check the results and report back.
October 6, 2026
Where you’ll meet agents. AI makers’ agent features, as each describes them on its own pages, with the date we read them. Alphabetical; not a ranking. Prices and features change often: check the maker’s page before you buy.

These products change quickly. Names, plans and prices in this section will date; the table is refreshed every quarter.

What it costs

A team of agents does more work than one chat, so it uses more computing, and on usage-based plans you pay for that. Anthropic reported that in its data “agents typically use about 4× more tokens than chat interactions, and multi-agent systems use about 15× more tokens than chats” (Anthropic). On a flat monthly plan the cost may show up as usage limits instead, or the feature may be reserved for a higher-priced plan.

There are other costs too:

  • More ways to go wrong. Each agent can make mistakes, and the lead can combine them badly. A reviewer agent helps, but it is still AI checking AI.
  • Harder to follow. With several agents working at once, it is harder to see why the final answer says what it does. Ask for sources and check them.
  • Time. Teams of agents can take minutes rather than seconds.

When it is worth it

Anthropic’s own write-up is a fair guide. It says multi-agent systems “excel at valuable tasks that involve heavy parallelization, information that exceeds single context windows, and interfacing with numerous complex tools,” and that tasks which “require all agents to share the same context or involve many dependencies between agents are not a good fit for multi-agent systems today” (Anthropic).

In small-business terms:

Worth considering for:

  • broad research, such as comparing many suppliers, competitors or regulations at once;
  • big jobs with separate parts, such as preparing a tender response with pricing, method and references;
  • checking important work, where having separate agents produce and review an answer catches more mistakes.

Not worth it for:

  • simple tasks a single assistant handles well, such as a reply, a summary or a rota;
  • tasks where each step depends on the last, where splitting the work only adds confusion;
  • anything where the extra cost and time outweigh the gain.

Your role: the manager

With an AI team, your job is the one a good manager does with a human team:

  • Set the goal clearly. What you want, for whom, by when, in what form.
  • Set the limits. What sources the agents may use, what they may not do, how much they may spend, when they must stop and ask.
  • Sign off. Read the result, check the sources behind anything that matters, and decide what leaves the business. That sign-off is the amber box at the foot of the chart, and it is never delegated to another agent.

Should a small business build one?

Rarely, for now. For most businesses, AI teams will arrive inside products: a research feature in an assistant, a help desk that uses several agents behind one chat window, an accounting package that drafts and checks. What matters is knowing what is happening behind the chat box, so you can judge the cost, ask the vendor how the work is checked, and keep your own sign-off. If you do want something built, start with one agent (Lesson 7) and add a second only when you can say exactly what it would do better.

Key terms

Every term is explained in the glossary.

Try this

Think of the biggest research or preparation job your business does a few times a year, such as a tender, an annual review of suppliers, or a plan for a new location. Sketch it as an org chart like the one above: the parts a lead agent could hand out, what each specialist would do, what the reviewer would check, and what you would sign off. Then ask: would one good assistant, given the same brief, do nearly as well? If yes, use one.

Sources