Default inference

When an agent needs to run inference and nothing more specific applies, it uses the default inference connection. Which connection that is, is a SuperAdmin platform setting — deliberately not an environment variable.

Setting it

A super_admin designates default_inference_connection_id on the Settings page — pointing it at one of the workspace's connections (a category='llm' connection). Leaving it empty means unset.

Recall the resolution precedence from Connections:

user → workspace → system default

The default inference connection is that system default — the backstop when a user or workspace hasn't bound their own key.

Strict mode

The companion setting inference_strict decides what happens when the default is broken or unresolvable:

• inference_strict = true (the default) — a broken default raises a visible error rather than silently falling back to some model default. You find out immediately that inference isn't configured correctly.
• inference_strict = false — restores graceful fallback to the model default when the designated connection can't be resolved.

Leave strict on

Strict mode is on by default for a reason: silent fallback hides misconfiguration, and in a sovereign posture you never want inference quietly routing somewhere you didn't intend. Turn it off only when you deliberately want best-effort behaviour.

Why it's a setting, not an env var

Keeping default inference in AdminSettings means a super_admin can repoint the platform's inference — say, to a different EU-ZDR connection — from the console, live, with the change audited, and without a redeploy. It's the same config-not-code principle that runs through platform defaults.