datasette-agent 0.4a0

Simon Willison has pushed out a fresh alpha of datasette-agent, version 0.4a0, and the headline feature is a new mechanism called await context.browser_task(). It sounds like plumbing, and in a sense it is, but plumbing like this tends to unlock a lot of downstream creativity.

What it actually does is let agent tools execute code directly inside the user's browser rather than staying confined to whatever the server-side Python process can see and do. Up to now, a Datasette Agent tool could query a database, shuffle some data around, maybe hit an API. But it couldn't reach into the page itself. With browser_task(), a plugin can run custom JavaScript on the client side, meaning tools can now inspect the DOM, trigger UI interactions, or read whatever state exists in the browser that the server never had access to.

Willison frames this as opening the door for a new class of Datasette Agent plugins. Instead of tools being limited to backend logic, plugin authors can build ones that reach across that browser boundary and do things that were previously awkward or outright impossible. That's a meaningful shift for anyone building agents on top of Datasette, since the browser is often where the actual user context lives — clicks, selections, visible tables, whatever's rendered on screen.

This is still labeled 0.4a0, an alpha release, so the API is presumably still settling. But the direction is clear: Datasette Agent is pushing toward tighter integration between server-side tool execution and whatever's happening live in the user's browser tab. Issue #33 on the project tracker has the technical details for anyone who wants to see how the mechanism is wired up.