Personal Agent Runtime
A model can propose an action without being the authority that permits it. This earlier engineering exploration puts a registered capability and policy boundary between a proposed plan and its execution, then checks the outcome.
The architecture separates planning, permission, execution, verification, and bounded recovery. Its offline acceptance path uses a fake client; this is software evidence rather than a demonstrated live-network run.
- Model proposes a plan
- Runtime checks capability and policy
- Execute within the boundary
- Verify, then finish or recover