SaaS consoles: let them bring their own agent

A user may want their own assistant to complete a repeated task in your console. Pick one task that the user can describe precisely, then decide which operations an assistant would need.

A browser tool can use the visitor's active session when the application registers and authorizes that tool. The public scan can inspect the entry page; it cannot verify the signed-in operation.

The economics

An in-app assistant and browser tools solve different integration needs. Compare them for the selected task, including authorization, model costs, client support and ongoing maintenance.

What to expose first

Look at what your users do repeatedly through the UI that the UI is bad at: anything requiring six clicks to answer one question, anything they export to a spreadsheet, anything they ask support to run for them.

query_metrics(metric, range, breakdown)   // readOnlyHint
list_segments() / create_segment(definition)
export_report(format, range)              // readOnlyHint
list_campaigns()                          // readOnlyHint
pause_campaign(id)                        // confirm before running
search_docs(query)                        // readOnlyHint

Start with a narrowly scoped read-only task. Review access controls and the data returned by each tool; readOnlyHint is an annotation, not an authorization check.

The refactor you will hit

A tool handler cannot reach into React state or a store the way a click handler can. In a mature console, a surprising amount of business logic lives inside components — and that is where the real effort in this project sits.

The move is to extract each operation into a plain function that takes arguments and returns data, then have both the click handler and the tool call it. It is a good refactor to have done regardless. It is also worth scoping before you promise a date, because add some tools and disentangle four years of logic from the view layer are different projects.

Consoles are where errors matter most

An agent may retry after an error. A useful failure message, such as `date_range` must be an ISO 8601 interval, gives it a way to correct the request; test the actual client behavior.

Write failure messages as instructions to a competent stranger. That is exactly the audience.