Choose the AI runtime

Settings → AI Configuration → AI Runtime answers one question for the whole workspace: who produces your completions. It needs the manage_model_policy permission.

The two backends

Agent runner is the default. Completions come from a connected runner driving a coding agent on a machine you control. Your prompts and your code stay on that machine, and the agent’s own subscription pays for the model.

Direct LLM is the fallback. Completions come straight from the model provider’s API using a credential you have configured. No runner is needed.

The distinction is not about cost. It is about where the work happens and whose credential pays for it, which is why it is a workspace decision rather than a per-agent one.

The default chain

Under the backend choice is the default runtime chain: the ordered list of runtime targets an agent inherits when it has no runtime of its own. The first entry is the primary; the rest are fallbacks tried in order.

A chain is worth having because runtimes fail for ordinary reasons. A machine sleeps, a subscription hits a limit, a CLI needs re-authentication. A chain means the next attempt uses the next target rather than the run stopping.

The flip

You do not have to watch this page. When a run has no usable runner, the platform falls back rather than failing: the next target in the chain is tried, and if the chain is exhausted the run reports what it tried and why each one was unavailable. The setting here is which order that happens in, not whether it happens.

A runner backend with no runner fails loudly

Setting the backend to Agent runner on a workspace that has never paired one makes every completion fail with a clear error rather than silently reaching for an API key. That is deliberate: a silent switch to a paid API is the surprise nobody wants. Pair a runner first.

When to override it

Per-agent, on the agent’s Runtime tab. Override when one agent needs something the rest do not: a coding agent that has to run in a specific checkout, or a research agent whose provider handles long context better.

Two things live elsewhere and are easy to confuse with this page:

• Which models are allowed at all is the model catalog.
• Whose runner executes a run is the execution mode on budgets. That is a different axis from who produces completions, and a workspace can sit on runner completions with managed execution or the reverse.

Next steps

• Execution backends: the concept behind the choice.
• The model catalog: which models the chosen runtime may use.
• Connect an AI credential: what Direct LLM needs before it works.