From sandbox to production: graduating an agent
The sandbox teaches you what the agent does. Promotion teaches you what it does at scale.
The public sandbox is where an idea proves itself: route a prompt, see the model choice, read the cost. But a sandbox run is anonymous, rate-limited, and disposable. Production is the opposite — owned, capped, and audited.
The gap you must close
A sandbox defaults wide: any prompt, any time, results shareable everywhere. Production defaults narrow: a scoped identity, a spend cap, an audit trail, and a lifecycle (active, paused, archived).
Promotion is the act of closing that gap deliberately:
- Provision a scoped agent with a namespace and API key.
- Set a realistic spend cap for real call volume.
- Pin a model route so a new model release does not silently change behavior.
- Point the audit stream at the owner.
Pause is a feature
The sandbox will tell you the agent is clever. Production will tell you whether it is safe. When it is not, paused stops the agent without deleting it — you keep the audit history and the config, and you can restore in one click when the fix lands.