A checkout dashboard goes from 87% completion to 75%. Someone needs to find out whether the page broke, a campaign brought in different customers, or the analytics changed. Google Stitch Loop can take that question to a connected codebase, suggest an explanation and prepare a fix. The most valuable response may be to leave the code alone.
In its November 18 launch, Google Labs describes a long-running product agent for software that already has users. You give it a standing objective, connect GitHub and supply the evidence it needs. Two decisions stay with the team. Someone chooses which proposed solution is worth building, and someone reviews the resulting pull request before it merges. As of November 19, Google is accepting waitlist registrations. Google Labs’ Stitch Loop announcement
- 01It watches what you connectA repository supplies the implementation. Logs, behavior data, feedback and product documents can help explain what users experience.
- 02It proposes a direction before writing the fixSuggestions cite evidence. Once you choose a solution, Loop writes code in an isolated cloud sandbox and opens a GitHub pull request.
- 03The measurement still needs an ownerA change in traffic mix alone can cause a 12-point checkout decline. Define the denominator, the segments and the outcome before choosing a fix.
- 04Access starts with a global waitlistAnyone can join the list. Eligible Google AI Ultra users receive priority. Google’s launch gives no general-release date or per-loop price.
01 — InputsGive the agent a question it can answer
A standing objective gives each loop something to look for on successive visits. Google lists logs, user behavior and feedback as possible inputs, alongside GitHub, and describes more than 180 integrations, custom Model Context Protocol (MCP) servers and local command-line uploads. What the agent can say about a particular checkout depends on the information the team actually supplies. A repository can show that an error message exists. It cannot establish how often buyers see it.
Google’s published command-line guide makes that boundary concrete. A workspace holds its attached repositories, objectives and context. An insight points to a route, and a solution proposes a change for that insight. Setup includes Google and GitHub sign-in and granting the Stitch GitHub App repository access. The guide documents uploads for individual files and code archives. Files supplied this way go into the workspace. Google’s Stitch CLI 0.13.0 documentation
For a checkout pilot, start with the storefront repository, a short definition of the checkout events and an aggregate count by route, release and device. Leave individual order records out if those counts answer the question. Add information only when a reviewer can explain what ambiguity it resolves. A support summary such as “keyboard users cannot move past the address field” answers a different question from a chart of total exits.
02 — WorkflowFollow the evidence to a pull request
Google calls the four stages suggestions, solutions, build and reflect. First, Loop identifies an opportunity connected to the goal and presents supporting product evidence. Next, it offers one or more solutions, with scope, tradeoffs and design options when relevant. This is the point to question the diagnosis. The CLI also lets a team dismiss an insight, a sensible outcome when a report rests on a renamed event.
Choosing a direction starts the build. Google says Loop writes code in an isolated cloud sandbox and submits a GitHub pull request for approval. Where they help explain the work, the request includes screenshots or video. A screenshot might show that an error is legible on a phone. A reviewer still needs to confirm that an actual user can recover from the error and finish the transaction.
After a change reaches users, Loop can follow positive and negative results in the connected data. It continues watching even after the objective is met. For this to be informative, give the system a way to associate a measurement with the version a visitor saw. A pull request that merged on Monday but reached only internal testers until Thursday has two very different dates attached to it.
03 — Starting pointMatch a first objective to the evidence
The three built-in loops cover latency, accessibility and cybersecurity. Teams can create custom objectives, including checkout conversion. Google says the security loop pairs Gemini 4 Carbon with Google DeepMind’s CodeMender and submits proposed fixes for review. A team already recording page performance by route may get a clearer first result from latency, because it has a baseline and can compare the same operation after release.
Choose an objective whose failure someone can verify independently. For a keyboard navigation issue, tell the reviewer which control must receive focus after an error and how to test it. For a latency target, specify the percentile, route and device class and count the affected visits. Our guide to browser rules for coding agents covers how to give an agent concrete web requirements.
04 — Worked exampleA 12-point fall with no regression
Suppose a store counts each checkout start once and treats a confirmed payment as completion. In the first period, 100 mobile customers start and 60 pay. Of 900 desktop customers, 810 pay. The store records 870 payments from 1,000 starts, or 87%. In the next period, the traffic mix shifts to 500 starts on each device. Mobile contributes 300 payments and desktop contributes 450. The total falls to 750 of 1,000, or 75%.
Mobile completion held at 60% and desktop at 90% in both periods. That 12-percentage-point drop is real for the business, yet a checkout patch has no identified regression to fix. If an agent suggested undoing the latest release, the team could dismiss that suggestion and open a narrower investigation into why mobile has consistently lagged. The figures here are illustrative, with no measured Stitch Loop result implied.
Now suppose mobile customers report a specific failure after address validation. Ask for evidence linking the complaint to a route, release and device. Define success as more confirmed payments per eligible mobile checkout, and keep duplicate charges and payment errors as guardrails. A cosmetic change that encourages more taps on the payment button could inflate click counts while making customers worse off. The guide to agent quality metrics explains a related measurement problem when counting an agent’s completed tasks.
05 — ApprovalWhat the reviewer approves and what ships
Google’s stated merge rule requires human review. Its launch describes an isolated cloud sandbox, but gives no technical commitment there about network egress, retained build artifacts or organization-specific isolation. Teams with requirements in those areas can use a small non-sensitive repository for evaluation while they establish the answers for their intended deployment.
The pull request is the handoff to the repository owner. Keep the usual tests, designated reviewers and release rules. A team might stage a change behind a flag and expose it to a fraction of visitors, or its existing pipeline might deploy every approved merge. Document which applies before granting repository access. The launch assigns Loop the code proposal and follow-up observation. It identifies no separate Loop-owned production deployment or rollback facility.
For the checkout example, the review packet should show the failing interaction, how the proposed change restores it, and which release and metric will verify the result. If a local reproduction matters, the Next.js agent devtools guide explains where a development server can help. Publish the release identifier alongside the measurement so someone investigating next month can find the change that actually reached customers.
06 — HandoffsWhere the product context travels
A design export and a Loop change serve different people. Google’s open-source Stitch SDK can extract HTML and screenshots from generated design screens. Those artifacts are useful to a designer or frontend developer. For a tracked change to a live application, Loop’s documented output is a pull request in the connected repository, sometimes with visual evidence attached. An exported screen alone says nothing about which code a production visitor receives. Google Labs’ Stitch SDK
Google also describes using Loop from Slack or as a source of persistent product context for coding agents such as Antigravity. The Stitch CLI and MCP make that context available to another agent. In Slack, the documented uses for @stitch include daily insight digests and creating loops. Daily refers to the digest. Google gives no fixed interval between samples of a production metric.
If another agent can consume the same context, decide which repository each agent may change and who owns the resulting requests. A developer responding to a digest should be able to find the original insight and discover whether a fix is already in review. Our guide to managing agents on GitHub covers how to keep those reviews visible when more than one agent contributes.
07 — AvailabilityWhat can a team try today?
Google Labs opened its global waitlist on November 18. The announcement extends registration to all users and gives eligible Google AI Ultra users priority. Joining the list establishes interest. It gives a team neither an invitation date nor a production launch commitment. The public announcement also supplies no per-loop price or service-level terms. Official Stitch Loop waitlist
An admitted team can evaluate one question before committing engineering time to more connections. Keep a record of the original observation, the proposed cause, the reviewer’s decision and the post-release result. Count both the time saved finding a problem and the time spent dismissing unsupported suggestions. For teams whose events change from release to release, fixing measurement may be a better first project than adding another objective. Our web development team works on the instrumentation and application changes behind those decisions.
Start with one decision you can verify
Pick a route with a known owner, a trustworthy measurement and a practical way to reverse a change. Ask whether the agent can connect a user-visible problem to the right code and whether the reviewer can tell when the suggestion is wrong. In the checkout example, recognizing the change in customers was already a good result. It saved a team from shipping an unnecessary fix.