Choose a model
In command examples, replace AGENT with your CoAgent's username or CoAgent ID.
The CoAgent's runtime supplies its available model catalog. Availability depends on the connected provider account and runtime.
Discover models
Use the AI provider controls on the CoAgent's detail page, or run:
socra agent model list AGENT
socra agent model refresh AGENT
Listing uses the saved catalog. With a shared connection, refreshing queries the provider through that connection without waking the CoAgent's computer. For a Codex CoAgent using a local sign-in, refreshing wakes the computer if needed. Resume a paused CoAgent before using that local path. Connect the provider first if the catalog is unavailable.
Save a selection
To configure the model in Admin:
- Open the CoAgent's detail page in Admin.
- Under AI, select the Model row.
- If the catalog is missing, select Load available models. To update an existing catalog, select Refresh models.
- In Model, select an available model.
- In Reasoning effort, select a supported value or the model's default.
- Select Save changes.
For the equivalent CLI operation, run:
socra agent update AGENT --model MODEL --reasoning-effort EFFORT
Replace MODEL with a model from socra agent model list AGENT and EFFORT with a reasoning effort that model supports.
Use --reasoning-effort default to use the selected model's default. New Codex CoAgents default to gpt-6-sol.
When changes take effect
For Codex, saved settings apply when a new turn starts. A notification steered into an active turn keeps that turn's existing model.
For Claude Code, saved settings apply when a new runtime session starts. Further messages can continue the existing session with its previous model and reasoning effort. A new computer session, after the CoAgent sleeps and starts again, loads the saved settings.
Changing a setting does not interrupt current work. Wait for the CoAgent to finish before arranging a restart.
Confirm the setting
socra agent get AGENT
Confirm that model and reasoning_effort match your selection. This reports the configured model, not independent proof of the model that served a particular past turn. A CoAgent's verbal description of its model can be stale; runtime turn metadata is the evidence for historical execution.
See Troubleshooting for related guidance.