This guide covers Agents → Custom code, where you import and deploy your own repository. For dashboard-built agents with uploaded knowledge, simple schedules, and scoped app keys, start with the Agents quickstart.
Choosing the agent runtime
Pick the runtime when you create the project: Web App keeps a container always on; AI Agent runs your agent serverless instead. From the terminal, deploy straight from a repo URL:
setcloud projects deploy https://github.com/you/support-agent --agent--agent is shorthand for --runtime serverless; leave both off and the project deploys as a container. Every push after the first build rebuilds and reships automatically, same as a Web App.
Cold starts and always-warm
Between requests an agent scales all the way to zero — no active agent compute while idle; storage and other linked services can still incur charges. A new request starts an execution; startup latency depends on the runtime and workload. If that wake time matters for your use case, turn on the always-warm toggle on the project’s page: it keeps one instance ready to reduce cold-start delays, billed as a flat reservation for as long as it’s on. Always-warm still respects your spend cap — hitting the cap pauses the agent, warm reservation included.
How requests are metered
Usage draws from the same credit balance as every other product: metered per request plus the per-second compute the agent actually uses while handling it, recorded in usage and deducted through billing. Model usage is charged separately when performed. Idle execution has no runtime charge when always-warm is off; retained resources can still bill. Rates are on /pricing.
Pooled session state, schedules, code-review tasks, and task runs are covered in agent schedules, tasks, and durable sessions. For build/runtime log locations and failed-deploy troubleshooting, see deployment logs and troubleshooting. Product overview: /products/compute. Rates: /pricing.