Orchestration

When a solution has more than one agent, orchestration decides who answers. The setting lives in Settings → Orchestration.

The two modes

Mode	Behaviour
Single Agent	Only the primary agent answers. Simple, predictable and cheaper.
Multi-Agent	MarIA picks, on every turn, one or more agents based on their descriptions.

Always start with Single Agent. Only move to Multi-Agent when you genuinely have agents with distinct specialities — support and billing, for example.

Primary agent

The agent that answers by default and coordinates the others. Pick the most general one: it is the one that takes the message when no other is clearly a better fit.

Max Participant Runs per Turn

Limits how many agents can be engaged on the same message. A high number makes every turn more expensive and rarely improves the answer — watch the effect on the consumption dashboard.

Trace narration

Provider, Model and Trace Narration Prompt define how MarIA describes, in plain language, what happened during the turn — which agents took part and which tools ran. It is the text that shows up in the Activity block inside the chat.

Use an economical model here: narration is a reading aid, not the answer to the user.

What actually drives the choice

Not the orchestration settings: the description of each agent. MarIA compares the message with the descriptions and picks. So:

• write short, concrete, mutually exclusive descriptions;
• avoid generic descriptions such as “answers customer questions” on more than one agent;
• if an agent is never picked, the problem is almost always its description — or it is inactive.