Switch agents and models
One Admiral session can run on different agents and models over its life. You switch from the composer picker; the transcript stays.
What switching means
Two cases behave differently:
Same agent, different model. The session resumes normally with the new model. Nothing is lost.
Different agent. A switch to a different agent starts a fresh native session under the hood. Admiral hands the new agent a compact summary of verified project facts so it can continue usefully. The visible transcript stays in place.
What does not carry across an agent switch:
- The agent's hidden reasoning.
- Tool state inside the old agent.
- Anything the old agent knew but never wrote into the transcript.
The new agent knows what the transcript shows, plus the handoff summary. When a switch loses something important, paste the missing context into the composer.
Which agent runs which model
Admiral routes your model choice to the right agent:
| You pick | It runs on |
|---|---|
| A Claude model | Claude Code, natively. |
| A Gemini model | Antigravity, natively, shown as Gemini. |
| Any other model | OpenCode. |
Claude or GPT models reached through OpenCode are OpenCode sessions. They are not Claude Code sessions, and their tools and sign-ins can differ.
Enable agents and manage sign-ins
The Providers settings page owns this:
- Sign in to each provider from its card.
- Turn agents on or off for the pickers.
- Watch usage meters, which warn at 80 percent and again at 95 percent of a limit, with the reset time.
See Settings pages.
Legacy agents
Codex, Cursor, Grok, Hermes, Pi, and Devin exist behind the legacy agents switch in Settings. Legacy agents accept new instructions at turn boundaries, so the composer offers Queue rather than Send. Their old sessions stay readable.
Next steps
- Agents and models: the concepts behind this guide.
- Steer and queue: why legacy sessions queue.