The authority problem in autonomous AI
Autonomous AI agents take actions — they call APIs, modify databases, approve transactions, deploy code, and interact with external systems. Each action implies authority: the right to take that action, on behalf of some principal, within some scope. Without explicit authority management, the agent operates on implicit, open-ended authority — it can do whatever it determines is appropriate.
This is the governance gap that kill switch architecture addresses. An agent with open-ended authority cannot be meaningfully controlled, because there is no boundary to enforce. Bounded authority creates the boundary: the agent can do X but not Y, up to Z threshold, within W time window, on behalf of P principal.
Components of bounded authority
A complete bounded authority model includes: (1) action scope — which actions the agent can take; (2) resource scope — which data, systems, and environments the agent can access; (3) value thresholds — maximum transaction amounts, modification sizes, or impact levels; (4) temporal bounds — when the authority is valid and when it expires; (5) delegation limits — whether and how the agent can delegate authority to sub-agents; and (6) jurisdiction — which regulatory or organizational boundaries apply.
Each component is enforced by the control plane, not by the agent. The agent proposes an action; the control plane evaluates it against the authority grant and returns a decision. The agent cannot override, bypass, or modify its own authority.
Delegation chains and traceability
When an agent delegates authority to a sub-agent, the delegation must narrow the authority — the sub-agent receives a subset of the delegating agent's authority, never more. The delegation chain must be traceable: any action taken by any agent in the chain can be traced back to the human principal who originated the authority.
Traceability requires that each delegation event is recorded in the evidence trail: who delegated, to whom, what authority was granted, what constraints were applied, and when the delegation expires. The chain is only as strong as its weakest link — if any delegation event is unrecorded, the traceability guarantee breaks.
Authority in DecisionHypervisor
DecisionHypervisor implements bounded authority through authority grants, delegation chains, jurisdiction scoping, and time-limited execution tokens. Every proposed action is evaluated against the actor's authority grant: is this action within scope? Is the resource accessible? Is the value within threshold? Is the authority still valid? Has the delegation chain been properly recorded?
Actions that exceed authority are denied or escalated for human review. The authority model is independent of the AI system — it is defined by the organization, enforced by the control plane, and recorded in the tamper-evident evidence trail. Revocation is immediate and propagates across all systems that recognize the authority.