Language and display
BE THE BOSS OF AI.ADRIAN DUNKLEY ↗
← This newspaper edition

Google’s Stitch Loop proposes website fixes for a human to approve

The Google Labs experiment is taking waitlist applications. Its proposed repair process keeps a reviewer between an AI-written change and a live site.

A navy loop surrounds sandstone discs on cream paper, with a vermilion block beside a gap in the loop.
The Unfinished Circuit · AI-generated artwork · AI Boss Ltd

Google has announced an AI tool intended to keep improving software after launch. For a Caribbean business with a small website team, the useful question is who checks a proposed repair before customers encounter it.

Stitch Loop, announced by Google Labs on 18 November, connects software code with product information such as error records and feedback. Google says it can suggest improvements, write changes in an isolated cloud workspace and submit them for approval. A pull request (PR) is a proposed code change that a reviewer can examine before accepting it. The announcement describes a waitlist, not general access or a service we have tested.

The owner still needs to define a good repair

Google describes default work on security, loading delays and accessibility. Those are provider-described capabilities, not independent findings about reliability. Google has opened a global waitlist, with priority for eligible Google AI Ultra users. That queue does not establish when a Jamaican applicant will receive service, at what price or on what terms. Google Labs product director Justin Toupin writes, “Nothing merges until you review the PR.” Here, merging means accepting the proposed change into the shared code.

A shop owner need not wait for access to improve the brief given to a developer. State the fault, the intended behaviour and what must remain intact. That also makes an eventual AI proposal easier to judge.

Try a booking-form brief on a test copy

Consider a hypothetical guesthouse form that requires a flight number before a visitor has booked a flight. Give the developer this brief: “On the test site, make flight number optional. Keep name, email and arrival date required. Keep the existing confirmation message. Do not change payment or cancellation terms.” Use the fictional name Test Visitor, email visitor@example.com and arrival date 1 December 2026.

Ask for a test submission without a flight number, a rejected submission without an email, and a phone-width screenshot. Check that someone using only the keyboard can reach every field and the submit control. Check the confirmation appears once after one submission. The test copy must avoid sending email or taking payments. A screenshot alone cannot establish those behaviours; the reviewer must try them.

Check whether anyone can supervise the agent

The AI Agent Readiness Score helps a website owner and technical lead identify gaps before giving software permission to act. It asks about documented processes, data and system access, oversight and the value case, then returns a score, a category and recommendations by section. An agent here means software that can take actions through connected tools.

Bring the booking process, access list and recovery plan to the questionnaire. Answer from those records, then compare the result’s sections. In our non-sensitive hypothetical test, selecting the second answer on each of its 12 questions produced 33/100 and “Foundations First”. That is an observed rule-based output from invented inputs, not a measured readiness level for any company.

If nobody can stop an unwanted change or restore the previous site, assign that responsibility and test recovery before connecting an agent. The score does not verify the controls or predict whether a pilot will succeed.

Sources and further reading

Sources checked 19 November 2026.