LLM settings
By the end of this page the router will use the model you chose, with routing decisions as steady as you want them.
Prerequisites
- A system created (Create a MAS).
- Admin or Developer role.
- Optional: a project LLM configuration — see LLM Settings for how these are defined.
What these settings control
These settings apply to the routing decision only — the short classification call that picks an agent. They have no effect on how the chosen agent answers; each member uses its own LLM settings for that.
Open the section
With the Config pill selected, expand LLM Configuration in the left panel.
Pick a configuration
LLM Configuration lists your project’s configurations, plus Use project default. Selecting one also adopts its model and temperature as the starting values.
Classification is a small, structured task — a fast, inexpensive model is usually the right choice. Save your strongest model for the agents that write the answers.
Set the temperature
Temperature accepts 0.0–1.0. Lower is more deterministic, which is what you want here: the same question should reach the same specialist every time. Raise it only if you have a reason to.
Set the token budget
Max tokens caps the classification response, between 100 and 2000. The router only needs to return a choice, a confidence score and a one-line reason, so a modest budget is enough. Raising it doesn’t improve routing.
Save
Click Save in the header. The router card’s LLM Config row shows the selected configuration, or Project default.
Cost and latency
Every classified turn makes one extra model call before the agent even starts. That’s the price of routing, and it’s why the shortcuts matter:
- Small talk is handled without a model call.
- A single-member system never classifies at all.
- A shorter Router Prompt and fewer few-shot examples mean a smaller, faster classification prompt.
Leaving Use project default selected is a fine long-term choice. Then changing the project’s default model changes routing too, with no edit here.
Common issues
- The selector is empty — The project has no LLM configurations yet. Use project default still works.
- Routing answers vary for the same question — Temperature is too high. Try
0.1or lower. - Routing feels slow — The configured model is a large one, or the classification prompt has grown. Pick a faster model and trim the prompt and examples.
- A deleted configuration was selected — Reselect a valid configuration, or switch back to Use project default, then save.
Next
- Fallback & handoff — handle the messages nothing matches.