As agents handle more of the routine work, the moments humans still touch software become higher-stakes. Approvals where your reputation is on the line. Overrides when a recommendation feels wrong. Exceptions the system flagged but couldn't handle on its own.

Those sessions demand sharper product design than anything most teams have built.

I wrote about five principles I keep coming back to. Three that feel most underbuilt:

Design living surfaces that assemble around the decision at hand. If the agent is 95% confident and uncertain about one parameter, show that one thing front and center.

Build for infrequent, high-stakes use. At twice a week, every friction point is a failure. Measure decision latency, override rate, and escalation precision. Session duration tells you nothing.

Design for progressive trust. A permission ladder from read-only to full orchestration. Each rung needs a different product surface. Every decision trace builds switching costs.

The other two: treating agents as first-class consumers and closing the feedback loop between agent behavior and human judgment. The full piece goes deeper on all five.

Full piece is on Substack. Link in comments.