Getting started
Start with a precise task, defined evidence, and approval boundaries—not an unrestricted agent prompt.
Private internal project. Features described here are evidence-limited and are not offered as a public execution service.
1. Define an auditable task
Select a specific code defect or change, a clean source revision, and the small set of files allowed to change. Define visible tests that can supply actionable diagnostics and separate withheld acceptance checks that the worker cannot inspect.
2. Preview execution scope
Review the requested model, budget, source and test identities, proposed capabilities, and expected outputs before permitting execution. Read-only inspection is distinct from sandboxed code generation. An agent cannot approve its own increased privileges.
3. Evaluate the candidate
An authorized test environment evaluates the proposal without applying it to the original source tree. Keep the candidate diff and its receipt together. Independently verify the signature, artifact hashes, test results, and scope.
4. Decide what happens next
Accept only after required tests and review. A rejected candidate should remain reviewable without changing the checked-out source. Avoid a blind retry where the worker has no new actionable feedback.
Availability
This is an explanatory workflow, not a public setup guide. Access to internal execution environments and grants is not offered by this documentation.