The Team Manager is the agent that decides which team member handles each request.
Overview
The Team Manager is the AI that runs a team — it reads each request and routes it to the member agent best suited to handle it. The screen — "Select AI models to manage and plan your team's work" — configures the manager with three things: the model it runs on, its instructions, and the team's conversation starters. There is no routing toggle or mode to pick; you define how the manager delegates entirely through its instructions.
Choosing the Manager's Model
Set the Team Manager LLM — "Select a LLM to act as your team's manager." This is the model that reads requests and decides which member to call, so pick a capable model that's good at understanding intent and following routing instructions. The manager only coordinates; the member agents do their own work with their own models.
Telling the Manager How to Route Work
Write the Team Manager Instructions — a free-text field where you describe how the manager should delegate. This is where routing lives: spell out which member handles which kind of request, how to split a multi-intent request across members, and any rules for when to do what. A typical instruction reads like "You are the team manager of multiple AI agents. Identify the core object of each query and route it to the agent best suited… Split multi-intent queries and route each separately…" Clear, specific instructions are what make the manager route well — there's no separate routing setting to fall back on.
Conversation Starters
Conversation starters are the suggested prompts users see when they begin talking to the team — the same idea as an agent's conversation starters. Add a few that show what the team can help with, so users have an easy first message.
What Tools the Manager Actually Works With
The manager doesn't get a normal agent toolset — it gets delegation tools built fresh each turn from the current eligible-member list:
A delegation tool whose member-name choices are constrained to the roster, taking the member's name and a written instruction (the "query"). In the standard setup it also takes a sub-task thread id, so a follow-up to the same member continues the same thread. The platform can also be configured to present each member as its own hand-off tool (named after the member and described by its profile) — same behavior, one tool per specialist.
The platform's standard default tools, resolved from the team's own settings, just as for an agent.
The manager is also instructed to reuse a sub-task thread for follow-ups to the same member on the same task, and to start a new thread only for a new sub-task or a different member.
What the Manager Writes When It Delegates
A detailed, self-contained instruction — the delegated member does not see the raw user conversation, only what the manager writes into the delegation (plus the earlier messages of that same sub-task thread). The delegation tool explicitly asks the manager for "all necessary context, constraints, and required output format so the agent can complete the task independently". This is why manager instructions that say "pass along all relevant details" materially improve results: anything the manager omits, the member never sees.
Does the Manager Answer the User Itself?
In the standard setup, yes: the member's answer is returned to the manager as a structured result, and the manager decides what to do next — invoke another member, ask a clarification, or compose the final reply the user sees. Teams can also run in platform-configured router or orchestrator modes: there the routed member's answer is published directly to the user, and the manager's job ends once it has routed (in orchestrator mode it is explicitly told the response was already published and to end the conversation).
Note: If a routed member fails, the manager receives a failure result telling it to retry, fall back to another member, or surface the problem — routing failures don't silently drop the request.
When a Member Comes Back to the Manager Instead of Finishing
When it needs something only the user can give. If a delegated member asks a clarification (or needs a connection chosen) while working a team sub-task, the question is handed back to the manager rather than posed to the user directly; the manager relays it, collects the answer, and re-invokes the member on the same sub-task thread so the member picks up where it left off.
The platform also protects the loop itself: a conversation-wide cap on agent invocations (60 by default, adjustable via the team's response-generation limit) stops runaway routing chains, repeated identical tool calls are detected and interrupted, and after repeated errors the manager is pushed to give its best final answer rather than loop forever.