Agent products need workflow state

The best AI products don't feel like chatbots. They feel like someone quietly cleaned up the workflow graph behind the scenes.

A chat box is a great demo and a poor product surface. It hides the two things that actually make an agent useful: the state it's operating on, and the actions it's allowed to take. When those are invisible, every interaction starts from zero and the user has to re-explain the world.

The interesting work is in the middle of that line, not the ends. Models are abundant. The scarce thing is a harness that holds context, exposes the right tools, tracks where a task is in its lifecycle, routes the uncertain cases to a human, and closes the loop with notifications and feedback.

I'll write the longer version of this argument in Agent Harnesses, Not Chatbots. This is the one-paragraph version: build the workflow, then let the agent move through it.