Every agent has a job, a scope and a reviewer.

One agent cannot check itself

Ask a single model to plan the work, produce the work and then confirm the work is sound, and what comes back is agreement, not verification. The reasoning under review is the same reasoning doing the reading. A pull request carrying its author's own approval was never reviewed, and that does not change when the author is a model.

Scopes and separation live outside the model

Each role's grant — repositories, Blueprint areas, capabilities, actions it must hand on — is held in configuration and checked by the control plane when the action runs, not asserted in prompt text. Two rules hold whatever you configure: the agent that raised a proposal is excluded from reviewing it, and critical work runs across more than one model provider. The platform refuses a run it cannot staff that way.

Delegation that leaves a record

You hand over the work of an engineering team without handing over control of it. Every proposal carries the role that raised it, the critic's findings, the policy result and the person who approved, so a change reads months later without replaying a conversation. Adding capacity to the portfolio then means adding roles, not extending trust.