Developers
Somewhere you cannot email a real customer.
Every developer gets an isolated environment with synthetic customers, fake invoices and mock tools. Nothing here reaches a production system, and there is no configuration that points a sandbox agent at a real capability.
Mocks that behave like the real thing
Including the awkward parts: rate limits, timeouts, partial failures and a customer with an open dispute. A mock that always succeeds teaches you nothing about your error handling.
Separate secrets, separate everything
Sandbox credentials reach sandbox systems. A leaked development secret reaches nothing that matters, which is the point of having two.
Seeded data you can rely on
The same synthetic customers and invoices every time, so a test that passed yesterday failing today means your code changed.
Promotion, not recreation
What you tested is what runs. Promotion moves a specific version between environments rather than rebuilding it, so behaviour cannot change on the way.
Where it stops
A sandbox proves your agent handles the cases the mocks represent. It cannot tell you how your real customers will reply, and a mock is somebody's model of a system rather than the system. For confidence about real behaviour, use a simulation against your own history — see /testing.
Developer platform
The rest of it.
Give AI a job.
Build an agent, choose what it can access, define what it may do, and deploy it to work safely across your business.
No code required · Full SDKs and API · Stop any agent instantly · [email protected]