AI/UXNov ’26Actively exploring

Leash

How far an agent may go on its own is a length of lead you hold, and you feel where it needs you.

Leash

Drag the grip to give the agent more lead. Where the line goes taut, it needs you.

Support agent · Tuesday, cheapest work firstOn its own 2 of 6
  1. Tag tickets40 by topic
  2. Merge tickets9 duplicates
  3. Send tracking12 customers
  4. Refund orders3 late · $86
  5. Issue a creditHalvorsen · $500
  6. Close accounts2 duplicates
A support agent’s morning, cheapest work first. Inside the lead it works on its own; the first step past it waits for you. Pull the lead back to halt a step mid-run.

How far should an agent get before it checks with you? Here the answer is a length of rope.

A support agent has six things to do this morning, cheapest first: tag tickets, merge duplicates, send tracking links, refund three late orders, put a $500 credit on an account, close two accounts. Drag the grip along the rail. Everything inside the lead runs on its own. The first step past it stops, the rope goes taut, and Approve appears.

Now pull the grip back while the refund is running. It stops and counts itself back down, one order at a time. Taking permission back should be at least as easy as giving it, and here it’s the same drag in the other direction.

Consentful Automation splits the same question up by capability. Leash puts it on one line, which only works because the plan is sorted by cost, and real work rarely sorts that cleanly. The rope is the part I’d keep.