The tool names below will change. This will not. An agent proposes a change, the twin is used
to test it before production sees it, and a human approves it by default. Every stage here is
planned, so read it as the shape of the argument rather than a path you can take today.
A gated agent write
Five stages, all planned. Approval is the default, not something you switch on.
1
Planned
An agent calls a write tool. The call is a proposal, not an instruction.
2
Planned
The call carries a tenant-bound credential. The tenant comes from the credential, never from the arguments, so an agent cannot reach another tenant by asking.
3
Planned
The change is tested in memory against the twin first, so its blast radius is known before anything touches production.
4
Planned
Approval is the default. Auto-approval is opt-in, per change type, and never the starting state.
5
Planned
The write records which agent made it and under which credential. Reads stay free; writes stay auditable.
An agent proposes a change and a human approves it by default. Auto-approval is opt-in, per change type.
Why Network changes are expensive to undo. Starting from human approval means trust is extended one change type at a time, by the people who own the outcome.