Runtime Mandate Check.
An AI agent is given bounded authority to spend. Before each consequential action executes, the proposed action is checked against the current human mandate.
The check returns one of four outcomes:
Runtime authority mandate demonstration, presented at MIT's Project NANDA hackathon (The Internet of AI Agents), 2026.
Five properties, made concrete.
- Bounded authority
- The agent can act independently within limits defined by a human principal.
- Action-time checking
- Authority is checked when intent becomes action, not only when access is initially granted.
- Human confirmation thresholds
- Some actions may be permitted only after explicit confirmation.
- Cumulative limits
- Authority may depend on prior actions and changing totals, not only the current request.
- Revocability
- The source of authority remains capable of changing or withdrawing the mandate.
Access is not authority.
An agent may possess valid credentials and access to an approved financial tool while still attempting an action outside the current intent of the person it represents.
A runtime mandate provides a distinct control point: before the consequential action happens, the system checks whether the agent still holds legitimate authority for that particular action, at that particular moment.
The scope of this demonstration.
This is a working demonstration of one authority-layer primitive. It is not a released commercial product and does not represent the complete OTIS architecture.
It demonstrates action-boundary mandate checking. It does not claim to prevent every cause of agent failure, inspect the agent's complete reasoning process or replace identity, security, sandboxing and monitoring systems.
The primitives, briefly.
- Principal
- The human or organization from which authority originates.
- Agent
- The delegated actor proposing an action.
- Mandate
- The live boundaries governing permitted action.
- Check
- Evaluation of the proposed action against the mandate.
- Outcome
- Approve, confirm, escalate or block.
- Evidence
- A record of the decision and its stated basis.
We are interested in real workflows where agents are expected to act on money, data, code, communications or operational systems.