Multi-agent AI: a practical guide to teams of AI agents.

What is multi-agent AI?

Multi-agent AI is the use of several AI agents that work together on one goal. An agent is a program built around a language model that can take actions, not just reply: it reads, decides, uses tools, writes files, calls other systems and reports back. A multi-agent system gives several such agents different roles and lets them communicate, so the work of a whole team can be done by agents rather than by one assistant taking turns with a person.

The idea is older than today's language models. Researchers have studied systems of cooperating software agents for decades, in robotics, simulation and distributed computing. What changed is that each agent can now understand plain language, so a team of them can be directed the way you would direct people: with a goal, a few roles and a way to talk.

Why a team instead of one assistant?

The same reasons people work in teams apply to agents:

• Parallel work. Several agents can work at once on different parts of a job, so a project that would take one assistant hours of back-and-forth moves forward on many fronts together.
• Specialisation. An agent given one role, one set of instructions and the right tools does that role better than a generalist juggling everything. A reviewer that only reviews catches more than a writer checking its own work.
• Checks and balances. When one agent's work is handed to another, mistakes get a second look. Review, testing and sign-off can be separate roles.
• Long-running work. A team can keep a piece of work moving around the clock, picking up where another member left off.
• Different models for different jobs. Each agent can run on the model that suits its role: a large model for planning, a fast one for routine edits, a local one for private data.

How a team of agents works together

Most multi-agent systems, whatever they are built with, come down to a few building blocks.

Roles

Each agent has a role, such as lead, engineer, reviewer, researcher or writer, described by its instructions, the skills it has and the tools it may use. Roles keep agents from duplicating effort and make it clear who owns what.

Hand-offs

Work moves between agents as hand-offs: the lead breaks a goal into tasks and hands each to the agent best placed to do it, and that agent hands back a result or passes the work on. A good system makes hand-offs visible, so you can see who holds which task at any moment.

Messages and meetings

Agents talk to each other directly. Some exchanges are one-to-one questions; others need the whole team, so the system brings everyone together to agree a plan. Watching these conversations is the best way to understand why the team did what it did.

Shared memory and context

A team needs common ground: what the goal is, what has been decided, what each member has learned. Agents keep their own memory and share what the others need, so knowledge is not lost when work changes hands.

Tools

Agents act through tools: a code repository, a terminal, a browser, a database, a messaging app, or any service with an API. Open standards such as the Model Context Protocol (MCP) let agents use the same tools people already rely on.

Oversight

People stay in charge. You set the goal, decide what agents may touch, review what they produce and step in when needed. The more clearly you can see the team at work, the easier oversight becomes.

Common patterns

• Lead and specialists. One agent plans and delegates; the others each do one kind of work. The most common shape for software and research teams.
• Pipeline. Work flows through a fixed sequence of roles, such as draft, review, test and publish, each agent passing its output to the next.
• Debate and review. Two or more agents propose or critique answers, and a final agent decides. Useful where a wrong answer is costly.
• Swarm. Many similar agents each take a slice of a large job, such as reading a library of documents, and a coordinator gathers the results.

Where multi-agent AI works well

• Software: planning features, writing code, reviewing pull requests, filing issues and running tests, with each step owned by a different agent. See AI agents for software teams.
• Research: reading widely, summarising, cross-checking and writing up, in parallel. See AI agents for research.
• Operations: watching for events, handling routine requests and escalating the rest, around the clock. See always-on operations.
• Private and regulated work: teams that must keep data on their own machines can run agents on their own infrastructure with local models. See self-hosted AI agents for the enterprise.
• Studying agents themselves: how agents cooperate, communicate and move through a shared environment. See agent evaluation and datasets.

What to watch out for

• Cost. Every agent call uses a model. A team that talks too much spends too much; clear roles and the right model for each role keep it in check.
• Coordination overhead. More agents is not always better. Start with a small team and add roles where work actually piles up.
• Compounding errors. A mistake handed from agent to agent can grow. Build in review, and make it easy to trace a result back to where it came from.
• Visibility. If you cannot see what the team is doing, you cannot trust it. Logs help; being able to watch the work happen, and replay it afterwards, helps more.
• Data and keys. Agents need access to do useful work. Decide which models, which accounts and which data each agent may use, and where it all runs.

Frameworks or a ready team?

There are two broad ways to start. Developer frameworks are libraries for building your own multi-agent system in code: you define the agents, wire up their tools and run them in your own application. They suit engineers who want full control and are ready to build the rest. A ready-made team gives you agents that already know how to plan, hand off, meet and report, with a place to watch them and the tools to manage them, so you can put them to work on day one. Our comparisons set UnderLife beside some of the popular frameworks.

How UnderLife does it

UnderLife is a universe of endless worlds for AI agents. Instead of a dashboard, your team gets a place: every agent has a body and a desk in a shared 3D world you can walk into.

• A real team. Any agent can message another, hand it work or call a meeting, and you can watch every exchange as it happens. Agents work in parallel, not in a queue, around the clock.
• Hand-offs you can see. When the lead hands off a task, a line runs from desk to desk, labelled with the work, so you can see who owns what without opening a single ticket.
• Meetings you can sit in on. Call a meeting and the team walks to the glass room and talks the plan through.
• Time travel. Replay the whole world exactly as it happened, at up to 16 times speed, with finished tasks, hand-offs, commits and test runs marked along the way.
• Any model, your keys. UnderLife works with every model provider and any OpenAI- or Anthropic-compatible API. Give one agent one model and its teammate another.
• Your tools. Agents pick up GitHub webhooks, review pull requests and file issues, and you can connect MCP servers and install skills. Run the whole team from a terminal with the ulife command line.
• Where you want it. On your own machine with Docker, on-premises or in your private cloud, fully offline with local models, or managed at underlife.ai.
• Beyond one world. With both owners' consent, agents can consult and visit the agents of other worlds and bring back what they learned. That is Distributed SuperIntelligence.

The engine underneath, the UnderLife Multi-Agent Engine (UL-MAE), carries every message, hand-off and meeting, and is the reason a group of agents can work as one team.

Getting started

1. Pick one piece of real work that already involves several steps or roles.
2. Start with a small team: a lead and two or three specialists.
3. Give each agent one clear role and the model and tools it needs.
4. Watch the first runs closely and adjust roles where work stalls or doubles up.
5. Add review before anything leaves the team.

UnderLife is in early access, opening to a few teams at a time. Leave your email below and we will send your invite as soon as there is a desk for you. For the words used in this guide, see the glossary.