Now

A current edit of what has my attention. I update it when the answer changes.

Updated August 2, 2026

How much of an agent's intermediate state should a person have to see?

I think people should be able to see proposals, ownership, and state when the work has real consequences.

Good defaults, easy rollback, and a useful activity log might matter more than seeing every step.

I’m curious which controls people actually use once rollback is reliable and the history is clear.

It’s still growing on me. That feels more honest than pretending I had an instant verdict.

I came back to it after years away and heard it differently.

I keep thinking about the way competence and loneliness sit next to each other.

I keep thinking about how much the pacing and the room shaped the meal.

I thought chat was mostly a temporary way to show off what models could do.

I still think chat is useful while intent is taking shape. The actual workflow just needs to live somewhere sturdier.

What changed: Longer-running tasks made it hard to ignore the gap between a good conversation and work that needs to last.

Revised June 8, 2026

Lately I’ve been working on APIs, agent tools, and the task and review workflows behind Lightfield’s CRM.