The Index
010

Field Note · 010

Trust, by drift.

On the small moment when the diff stops getting read. Spring 2026.

The first time an agent surprises a user is when it produces a draft worth sending. The second time, less remarked but more consequential, is when the user notices that the drafts have stopped being read before approval. Nobody told anyone to stop. The reading just thinned out.

01 /
The published research describes approval drift as a slow process measured across thousands of users. In practice it is faster. By the fifth successful action, the scan is down to the subject line. By the tenth, just the recipient. Somewhere after that, what is being confirmed is a green checkmark, not a message.
02 /
Trust between people is supposed to be slow. It is earned, lost, sometimes repaired, kept track of in ways neither party fully articulates. With an agent the timeline collapses. Five interactions and the user is in. There is no social cost to overriding it, which sounds like a feature. It is also the thing that removes friction from the calibration. Nothing pushes back on acceptance the way a colleague's hesitation would.
03 /
What is strange is the asymmetry. When the agent gets something wrong, the user notices and feels betrayed. When it gets something right, the user feels nothing. The emotional ledger always points away from trust. The behavioral ledger points the other way. The two readings disagree. The behavior is winning.

The agent is patient. It has been teaching us, one green checkmark at a time, to approve faster than we read. We have proven excellent students.

AI AgentsTrustDelegation