The Coordinator

The Coordinator is the agent at the front door. You talk to it, it figures out which specialist should answer, and either replies itself or hands the work off.

What it actually does

When you message the Coordinator, it:

  1. Reads your message.
  2. Looks at the agents you have, their roles, and their descriptions.
  3. Decides whether to answer directly or delegate.
  4. If delegating, opens a fresh ephemeral session with the chosen specialist and asks them.
  5. Collects the specialist’s reply and integrates it into the response back to you.

You can have multiple delegations chained — “Research the latest AI news and email it to me” might delegate to the Researcher first, then to Email.

Background worker teams

Ordinary prompts use one execution path, even when the work is divisible. A news request, comparison, audit, lookup, device action, or long job does not automatically create a worker or team. The Coordinator starts a detached worker only when you explicitly ask for background or agent execution. That worker stays single-threaded unless you also ask for multiple agents, parallelism, or fan-out.

If you give an explicit count — for example, “use 6 agents to audit these sources” — that count means six child workers; the detached coordinator is additional. If you explicitly request parallelism without a count, the coordinator chooses the smallest useful team, normally two to four children.

Requested team size and concurrency are separate. OpenEnsemble creates the full team plan and reserves every lane’s scope up front, then queues lanes that exceed the shared concurrency limit. The task card shows requested, active, queued, and completed workers. A user-specified team may contain up to 12 children by default; owners can lower that ceiling with OPENENSEMBLE_MAX_WORKSTREAMS_PER_OUTCOME. Requests above the configured ceiling are reported explicitly rather than silently reduced.

Choose worker models and effort

You can request a team on any models available to your profile, even when the parent uses another provider. For example: “Use two agents on my configured local model to compare these files,” or name a mixed team explicitly. OE checks each requested model and reasoning level; an unavailable choice is reported rather than silently replaced.

Requested team size counts child workers; concurrency determines how many run at once. Agent activity and Run Inspector show requested and resolved provider, model, and effort for each worker. Ask “check workers” or open the Agents pop-out for status and Stop controls.

In ensemble mode, a specialist can escalate a request to the coordinator when another capability is needed. Single assistant mode uses the primary’s enabled skills and private workers instead of named-agent delegation.

Direct chat vs. delegation

This is the most important Coordinator concept:

  • Direct chat with a specialist (clicking them in the agent list) → persistent session. Memory accumulates.
  • Coordinator delegation to a specialist (ask_agent) → ephemeral session. Every delegation is fresh, no carry-over.

So if you ask the Coordinator to “ask the researcher about X” twice, the researcher answers the second one with no memory of the first. To build context with a specialist, talk to them directly.

Tuning the Coordinator’s choices

The Coordinator picks specialists based on each agent’s:

  • Role (Coder vs. Researcher vs. Email)
  • Description (your one-line “what is this agent for?”)
  • The user’s message

If it’s picking the wrong one, the cheapest fix is to write clearer agent descriptions. (“image and video generation specialist” is more useful than “creative agent”.)

When to talk to a specialist directly

  • You’re iterating on a long task and want continuity (debugging a coder agent’s output, a long research thread).
  • The Coordinator keeps misrouting.
  • The specialist has its own UI or skill state you want to inspect.

When to talk to the Coordinator

  • Casual / mixed work where you don’t know who should handle it.
  • Multi-step asks that span specialists.
  • Telegram and other inbound channels — they always go to the Coordinator first.

No Coordinator?

You don’t have to have one. If you only have specialists, just talk to them directly. A Coordinator just makes life easier when you have several specialists.

This site uses Just the Docs, a documentation theme for Jekyll.