Skip to content

3.4. Agent Model Selection

Every agent run needs a model. Backend.AI GO resolves which model to use through a defined fallback chain, so an agent without an explicit model still runs as long as a sensible default exists.

Resolution order

When an agent run starts, the model is resolved in this order. The first source that yields a usable model wins:

  1. Explicit model — a model id passed directly to the run (for example, the model selected for the current chat, or a model supplied programmatically).
  2. Per-agent preferred model — the model set on the agent's profile (Preferred model in the Agent Marketplace editor, or the model field in the Squad agent configurator).
  3. Global default agent model — the value of Settings → Models → Default agent model. This is the safety net used by any agent that has no preferred model on its profile.
  4. Error — if none of the above is set, the run fails with a configuration error that names both places you can configure a model: the per-agent preferred model and the Settings → Models tab.

Empty values and the literal placeholder default are skipped at every step, so a stale or blank setting never reaches the inference router.

Sub-agent delegation

When an agent delegates to a sub-agent (for example via an @mention), the sub-agent's model is resolved through the same idea with one extra step before the global default:

  1. The sub-agent profile's preferred model.
  2. The parent run's model (so a delegated agent inherits the conversation's model when it has no preference of its own).
  3. The request-level model the caller passed (typically the current chat model).
  4. The global default agent model from Settings → Models.

This means the global default benefits both top-level runs and delegated sub-agents.

Setting the global default

Open Settings → Models and choose a model under Default agent model. Any agent whose profile has no preferred model will use it automatically — no per-agent change required.

The model pickers in the Agent Marketplace editor, the Squad agent configurator, and the squad bulk model-change dialog all show a short hint and an Open settings link pointing to this setting, so the fallback behavior is discoverable wherever you pick an agent model.