Agentic Systems

Agentic AI systems do more than produce a single answer. They may set subgoals, make plans, call tools, read and write files, browse, send messages, or take actions over time. Transparency must therefore include goals, plans, permissions, tool calls, approvals, and rollback paths.

The more agency a system has, the more important it is to know what it was trying to do and what authority it had.

Set Approval Boundaries for a Travel-Booking Agent

This agent can search, plan, and book travel on its own. For each step in its plan, decide whether it needs human approval before proceeding. Match friction to consequence.

1

Search flights

reversible, no cost

2

Draft itinerary

reversible, no cost

3

Hold a seat (no charge)

low consequence

4

Purchase nonrefundable ticket

high consequence

5

Notify user of booking

reversible, no cost

The irreversible, costly action requires human approval before it happens.
Approval should be tied to consequence. Drafting an itinerary can be automatic. Purchasing a nonrefundable ticket should require clear confirmation.

Agent traces should record the user goal, system instructions, plan steps, tool schemas, tool arguments where safe, tool results, intermediate decisions, final output, and any human approvals. This trace helps reviewers understand whether a bad outcome came from a bad goal, bad plan, bad tool result, or bad permission boundary.

Approval should be tied to consequence. Drafting an itinerary can be automatic. Purchasing a nonrefundable ticket should require clearer confirmation.

Make Agency Legible

Before an agent acts, users should understand the goal, the next meaningful action, and whether the action can be undone.

Checkpoint

What should be logged for an agent that can call external tools?