An agent token lets Claude (or a script, or CI) do what you do here, with no browser: deploy your branches, wait for them, read the deploy log and each service's logs, test your slot's APIs, redeploy and release. It acts as you, on your slots only; it never takes over someone else's. It lasts 30 days and is shown once.
Then tell Claude: read $SWAN_DEPLOY/agent.md and deploy my branch. In swan-services the swan-deploy skill does this for you.