Skip to content
AX Framework
Pillar 05 of 06

Trust & Control

Can an agent operate within clear permissions and safety boundaries?

AX proposal

Agents act with scoped, visible, revocable authority, and people stay in control of consequential decisions.

When an agent acts on someone's behalf, three parties need to trust the arrangement: the person delegating, the product being used, and the agent itself. Each needs to know what is allowed, what is not, and who is accountable.

Trust & Control is about making authority explicit. Permissions are scoped to tasks rather than granted wholesale. Sensitive actions declare their sensitivity. Some actions require a human to confirm, and the product makes that requirement visible before the agent commits to a plan.

This pillar is where AX meets security, privacy, and product ethics. It is also where the most expensive failures happen, so it deserves the most deliberate design.

What good looks like

Signals

  • Permissions are scoped, time-bound, and inspectable by the delegating person.
  • Actions declare a risk level and whether they require human approval.
  • Agents identify themselves, and the product can distinguish agent traffic.
  • Spending, data sharing, and irreversible actions have explicit limits.
  • Every agent action is attributable in an audit log.

Anti-patterns

  • Agents operating with a human's full session credentials.
  • Approval requirements discovered only after an action fails.
  • No way for a person to see or revoke what an agent may do.
  • Sensitive actions indistinguishable from routine ones.

Evaluation checks

Use these questions when reviewing a task. They feed the AX evaluation.

  1. T1Can permissions be granted per capability rather than all-or-nothing?
  2. T2Do sensitive actions declare that they require human approval?
  3. T3Is every agent action recorded with the identity of the agent and the delegating user?
  4. T4Can a person revoke an agent's access immediately?