Choosing the model your agents use

Standards does not marry you to any AI provider. Models are declared in your workspace, and your agents draw on them.

This is not technical vanity. Models change price, performance and terms several times a year. A workspace locked to one provider takes every one of those changes.

How the model is chosen

Each agent can name the model it uses. Failing that, it takes the workspace default.

Some internal tasks have their own model, independent of the one used in the conversation. Summarising a long exchange, for instance, does not need the most expensive model: that is compression work, not reasoning.

What happens when things go wrong

Two very different situations, often confused.

The model has gone from the catalogue, because it was retired or renamed. Here, a replacement model is chosen automatically: an agent configured on a reference that no longer exists keeps working.

The provider is down. Here, there is no switchover. The same model is retried, with a growing delay between attempts, up to a ceiling.

Three criteria to choose on

Cost. Usage is metered. A powerful model on a simple task costs money without bringing anything.

Performance. Extracting an amount from a PDF and writing an analysis do not ask for the same thing.

Confidentiality. Some files must not leave a given perimeter. This is often the criterion that decides, not the other two.

What does not change

Switching model touches neither the tools your agents have, nor their permissions, nor their memory, nor the guardrails set on them.

That is what makes the change possible without rebuilding anything.