In AI → Models. The agent is ours; the language model is an interchangeable part.
This isn't a matter of vocabulary — it's what guarantees that switching suppliers doesn't change your agent: the instructions, the allowed tables, the tools, the authorizations, and the screens stay the same.
Leave it on Automatic#
The platform chooses by considering three things, in this order:
- What the task requires. An agent that uses tools isn't sent to a model that can't use tools — checked before the call.
- Available credential in this environment.
- Provider health. If one fails repeatedly, it leaves the queue for a while.
If the main one doesn't respond, it tries the next — and the switch shows up in the run, with the reason. It is never silent.
If you choose a fixed model, you own that choice: there is no automatic switch if it fails.
Three statuses — don't mix them up#
| Status | Means |
|---|---|
| Implemented | there is support in the code |
| Configured | there is a credential in this environment |
| Has responded here | it actually responded, in this environment |
Support that's ready but has no credential is not a "working provider." The screen shows the three separately precisely so no one confuses them.
Next to each provider, its current status appears: Responding, Unstable, Down, or Credential not configured — and, when the credential is missing, the screen names the missing environment variable instead of just saying it's not available.
The Models in the catalog section lists the models the product knows, by provider.
No credential appears on screen — not even masked.
Cost#
Each run records the provider, the model, the tokens, and the cost. When the provider doesn't report the cost, the field stays empty — never zero. Zero would claim "it didn't cost anything," and that would be a lie.
Monitoring is in Dashboard.