I Gave My AI Browser a Recovery Routine
My AI team and I repaired an assistant's visible browser connection and tested recovery from a stopped screen. The instructions now define what it can restart, what it must verify and when a failure needs a different response.
Today my AI team and I restored an assistant's visible browser connection. We also changed its operating instructions so that a stopped screen has a defined recovery path, rather than leaving the next browser task dependent on another manual repair.
The immediate problem was the display connection. That matters because a browser that cannot reach its screen has not yet established anything about the website it was meant to use. Treating those as the same failure would send the investigation in the wrong direction.
Recover the missing prerequisite
We started the assistant's Bot Screen and enabled on-demand startup. Today's work journal records a native-browser check that loaded a harmless test page and captured a screenshot. The browser was also checked after restarting a stopped screen.
That second check was important. Starting something once demonstrates a working state. Exercising the restart path tests the particular recovery we are putting into the instructions. No basket items were changed during these checks.
The new procedure is specific: check the assistant's own screen state, start it if stopped, retry the browser, and verify the live page and capture. This applies to work that is already authorised. Recovering the screen does not extend permission for what happens on the website.
Not every failure deserves another retry
A genuine website denial is a different case. So is a configuration or resource failure. The updated instructions keep those separate from the stopped-screen recovery path, rather than treating every browser error as a reason to restart and try again.
There is still an open behavioural check. We verified the screen and browser setup, including recovery after a stop. We have not yet observed the assistant following the new procedure during its next relevant browser task.
Today's result is therefore a repaired connection, an exercised restart path and an explicit routine to follow. The next real task will show whether that routine is used without another intervention.
Evidence note: this account follows today's scoped work journal. The recorded browser checks were not independently repeated for this post.