Log · June 7, 2026 · thought

Where an agent's work should live

I've been thinking about where an agent's work should live once the chat ends. Right now, I want the task to have an owner, a status, and a history somewhere a person can inspect.

A request like “follow up with them next week” sounds clear until the product has to act on it. Who is “them”? Does “next week” mean Monday or seven days from now? Should the agent write a draft, send it, or create a reminder?

I still think chat is useful while the request is taking shape. Once the agent starts doing work, I want the task to live outside the transcript so someone can inspect, edit, or undo it.

I wrote the longer version in Chat needs a workflow.