Datasette Agent 0.5a0 puts chat beside the database app

Datasette Agent 0.5a0 adds a panel that keeps chat beside a table, query or database app. Combined with datasette-apps, it lets users prompt an app into existence and continue editing it from the conversation sitting next to the result.

The panel joins the conversation to a live Datasette page

A chat could already call tools and return results. The new panel keeps both work surfaces visible at once: the conversation continues while a specific Datasette page remains beside it. Users can see the table, query or app they are discussing.

Simon Willison describes the panel as the most notable feature in version 0.5a0. A public demo is available at agent.datasette.io after signing in with GitHub. The release remains alpha software, so this is a fast-moving tool rather than a stable production interface.

Editing beside the preview shortens the feedback loop

The useful change is shared context, not another chat window. A user can inspect a result, request a modification and check it immediately. For internal dashboards and data exploration, that removes some copying of table names, URLs and error descriptions between tools.

This release addresses a different layer from datasette-apps 0.3a0, which Radar covered earlier. Apps added narrower permissions around stored queries. Agent 0.5a0 changes the workspace itself. The releases complement each other, but the panel is not an update to the security model.

A visible result can encourage approval before verification

The panel improves orientation, but it does not validate a query or explain the consequences of a write. A convincing chart can still use the wrong filter, and an app can have broader access than expected. Bringing chat and output together shortens the review loop without removing the need to inspect data.

The source provides no latency measurements, error rates or user study. Claims about a better workflow therefore rest on the interface design and public demo rather than operational evidence.

Long editing sessions and resumable work will determine the panel's value

Future versions need to show whether the panel preserves the right context across many changes, supports several open pages and can restore an earlier session. A clear link between each request, executed query and app change will also matter.

If a user can explain an altered chart or filter after an hour of work, the panel may become a real workspace. If context disappears with the chat, it will remain mostly an impressive demo.

Lilith's verdict

Datasette has seated the agent beside the screen the user is actually watching. Now it must ensure that an hour of edits leaves more than a polished app and a chat full of tracks nobody can connect.