Multi-Agent System overview

A Multi-Agent System (MAS) puts several AI Agents behind a single conversation. To the person chatting, it looks like one assistant that happens to be good at everything.

What differs between systems is how those agents work together. That is the system’s pattern, and you choose it when you create the system.

You build a MAS in Agentic Workspace → MAS Systems.

The three patterns

Pattern	The idea	Who decides what runs	Reach for it when
Conversation Router	One specialist answers each turn	The router, fresh every turn	Your agents cover separate domains, and one of them can answer on its own
Planner / Executor	A planner breaks the goal into steps, and the team runs them	The planner, at run time	The work needs several agents and you cannot know the steps in advance
Agent Graph	You draw the flow: agents, connections, conditions and loops	You, when you build it	You know the sequence, and you want it to run the same way every time

The three pattern cards on the Create Agent System form

The pattern is set when you create the system, and each one opens its own editor. Pick it with care — see Create a MAS.

Choosing between them

Ask what decides the order of work.

Nothing decides it, because only one agent answers. That is the Conversation Router. Each message goes to one specialist, and the next message is classified again.
The system should work it out. That is Planner / Executor. You supply a team; the planner writes a plan for each goal and the executor runs it.
You decide it, in advance. That is Agent Graph. You draw exactly which agent runs after which, with conditions on the connections and loops where work needs another pass.

A single AI Agent already picks between its own skills. Reach for a MAS when the specialists deserve to be separate agents — separate prompts, separate tools, separate owners — not merely separate skills.

Skills matter in all three

Whichever pattern you choose, the skills on your member agents do much of the work — but each pattern uses them differently:

Pattern	What skills decide
Conversation Router	Which agent answers each message. Routing is classification against every member’s skills, so vague or overlapping ones send messages to the wrong agent.
Planner / Executor	Which agents a plan needs. The planner reads every member’s skills to choose.
Agent Graph	Nothing about which agent runs — you decided that by drawing the flow. Each agent still uses its own skills while it runs.

Skills are edited on each agent, not in the system builder — see Skills.

💡

Before building a system, open each candidate agent and check that its skills have real descriptions and example queries. For the first two patterns, that work pays off directly as accuracy.

When to use a MAS

Use a MAS when	Use a single AI Agent when
Distinct domains need genuinely different instructions, tools, or data	One set of instructions covers everything
Different teams own different agents and ship on their own schedules	One team owns the whole assistant
You want each specialist to stay small and testable on its own	The skills are few and closely related
You already have working agents and want one front door for them	You are starting from scratch

What a system is made of

Every system, whatever its pattern, has the same parts:

Piece	What it is
Member agents	The specialists. Each must be an AI Agent with status Active.
The pattern’s own configuration	Router settings, planner settings, or the graph you drew.
A version history	Save writes a draft; publish turns a draft into a production version.

Task Agents cannot be members. If a system needs a deterministic pipeline, expose it as a tool on one of the member agents instead — see Tools.

Lifecycle

The path is the same for all three patterns, and only the middle step differs:

Create the system and choose its pattern — Create a MAS.
Author it, in that pattern’s editor:
Test it in the builder — Test the MAS.
Publish a version — Publish.
Deploy it — Deploy.

Availability

Starter MAS Systems is available on Starter and above. On the Free plan the section is hidden from the project sidebar. All three patterns come with it — there is no extra plan gate on any of them.
Member agents must be published before the system can be deployed.