Observation, policy decision, intervention, and evidence administration should be separable and independently auditable.
Why this question matters
Autonomous monitoring creates its own concentration of power. A guardian with universal visibility and blocking authority can protect the system, but compromise or bias turns it into a privileged control point.
Separation of duties can remain automated. One component can compute risk features, another can apply policy, and a narrowly scoped actuator can pause the relevant task. High-impact or ambiguous cases can require human approval without slowing routine low-risk decisions.
Signals worth observing
- One agent can modify policy, suppress alerts, and approve execution.
- Interventions leave no independent evidence trail.
- Guardian access exceeds the systems and data it monitors.
Practical control direction
- Use distinct identities and credentials for each guardian role.
- Write evidence to an append-only independent store.
- Require dual control for irreversible or organization-wide intervention.
AgentCollusion lensA trustworthy oversight layer applies the same relationship analysis to its own components.
