Give every action a clear basis, scope and responsibility. Explore the boundary between models and tools.
Separate planning from execution
Agents in real workflows need a separation between candidate actions and execution. Models interpret goals, organize information and propose plans. Policies, permissions and authorized people determine whether those plans may run.
From tasks to receipts
HIVM structures a task as a plan, then associates tool scopes, risk levels and authorization requirements. After validation, simulation and any required human approval, the executor processes the action and records its state and evidence.
Make exceptions part of the workflow
Timeouts, uncertain states and external failures cannot always be reduced to success or failure. The architecture needs reconciliation, duplicate prevention and pause, review or compensation paths. Human approval remains part of the path for high-risk actions.
Current research scope
The HIVM materials describe a layered runtime and security architecture. They define controlled execution between models and tools; product access will progress with development and validation.
