Retarget a running agent's sticky runtime settings.
Changed onSwitches the model a running agent uses from its next response on, the only runtime setting today. The new model is validated exactly as a launch with it would be (present in the agent type's catalog, enabled by admin policy, and reachable with the credentials the requesting user can see, on the run's own route). The switch is posted to the agent's box as the harness's own instruction, which the box keeps across suspends (a sleeping box wakes to take it), and then recorded on the agent, so it is what cost attribution records from here on. Returns the updated agent record. 409 agent_not_switchable for a run that is still starting or can no longer take a turn (stopping/stopped/error/terminated), 400 model_switch_unsupported for a harness that takes its model only at startup, 400 agent_type_has_no_model_catalog for a single-model agent type, 400 model_not_in_catalog for a model outside the catalog, 400 model_not_reachable_for_run for a model on the other route than the run's (direct or AI Gateway), 403 model_not_allowed when admin policy disables the model or its provider, and 503 model_switch_not_delivered when the instruction could not be posted, in which case nothing changed.
Request
- Base URL: /api (relative, and the document was published on a file host, not at the API’s origin)
- URL: /api/agents/{id}/runtime-settings
- Auth: none declared. Required headers: x-organization-id
Path parameters
Headers
Active organization id (or slug). Required for org-scoped operations; the generated client enforces it at compile time.
Request body
Response
Response
Changes
- info
endpoint added
- info