Simon Willison released datasette-agent 0.5a0, a versioned build that puts an AI assistant alongside a live data tool. The design splits the screen: one panel shows the agent's work, the other shows the database view the agent is reading and writing. A human can watch both sides at once and decide whether to keep going.
The split panel is the interaction. The human does not wait for a finished answer. They see the agent build a query, watch the result appear in the database view, and step in if something looks wrong. The agent's actions stay visible as they happen, not after. That changes what the human can do: they can redirect early, before a bad query runs twice.
For teams designing agent features inside data products, this is evidence that putting the agent's working view next to the product view does real work. The human stays oriented. The agent's choices stay legible. The gap between 'what the agent did' and 'what I can see' closes, and that is where most agent UX falls down.