Agent Permissions, Risk Limits and Human Escalation
How to apply least privilege, deterministic limits, approval gates, and escalation paths to autonomous Quantin agents.
Agent safety depends on the authority around the model. Prompts can guide behavior, but permissions and risk limits must be enforced by systems the agent cannot rewrite.
Apply least privilege
Start with no access and add only the capabilities required for the defined responsibility. Separate permissions to read, propose, approve, and execute. Scope credentials to specific environments, resources, and operations.
Use deterministic risk limits
Limits should be evaluated immediately before an action. Examples include maximum order size, daily action count, allowed destinations, permitted data classifications, working hours, and aggregate exposure. A model recommendation never overrides a failed policy check.
Design approval gates
An approval request should contain the proposed action, expected effect, supporting evidence, confidence, alternatives, and expiry time. The reviewer identity and decision must become part of the audit record.
Escalate clearly
An agent should escalate when required information is missing, sources disagree, a request exceeds permission, a policy threshold is reached, or an external system behaves unexpectedly. The escalation must state what is blocked and what decision is needed.
Example authority levels
| Level | Agent authority |
|---|---|
| Observe | Read approved sources and report findings. |
| Recommend | Prepare a structured action proposal. |
| Execute with approval | Act only after an authorized person approves. |
| Bounded execution | Act automatically within enforced limits. |
Review permissions continuously
Audit unused capabilities, denied actions, overrides, and credential age. Remove permissions that are no longer necessary. When the agent’s responsibility changes, treat it as a new authorization decision rather than silently expanding access.
The objective is useful autonomy with predictable failure behavior. Clear limits make an agent easier to trust, test, and operate.