Google Stitch Loop: what it is and how to get access
2026-11-22 · 4 min read
Google Stitch Loop is a new Google Labs experiment: an agent that connects to your code and your live product data, suggests improvements, builds them, and opens a pull request for you to review. You cannot use it freely yet. Google opened a waitlist on November 18 at stitch.google.com/loop/waitlist, and Google AI Ultra subscribers get priority.
What Google posted
The @stitchbygoogle account introduced it as "a product agent built for outcomes, not just PRs." The pitch is a split in duties: Stitch helps you design a product, and Stitch Loop keeps improving it once it is live.
The main @Google account made the same point from the other side. Coding agents stop when the code ships, it said, while Stitch Loop "keeps working after launch." That post also names the models behind it: Gemini 4 Carbon and DeepMind's CodeMender, the security agent.
How a loop works
Google's announcement, written by Justin Toupin of Google Labs, describes a loop as four stages that repeat:
| Stage | What Stitch Loop does |
|---|---|
| Suggestions | Finds opportunities in your product evidence: logs, user behavior, feedback |
| Solutions | Offers one or more fixes per suggestion, with tradeoffs and design options |
| Build | Writes the code in an isolated cloud sandbox and opens a GitHub pull request |
| Reflect | Watches the live numbers afterward, including after the goal is met |
It ships with three default loops: security, latency and accessibility. You can add your own, and Google's example is improving checkout conversion. The security loop is the one that uses Gemini 4 Carbon together with CodeMender.
If you own the product, the sentence to remember from the announcement is that nothing merges until you review the pull request. The agent proposes and builds, and a person still has to approve.
How to get access
- Join the waitlist at stitch.google.com/loop/waitlist. Google says it is open to all users globally.
- If you pay for Google AI Ultra, you are in the priority group. Google has not said when invites go out, and it has not published a price.
- Once you are in, connect a GitHub repository, then connect product context. Google says there are over 180 integrations, plus custom MCP servers and local uploads through a command line tool.
- Pick a default loop or write a custom one with a clear objective.
TestingCatalog got an early look and published a walkthrough on November 21. In its account, you create a workspace for a GitHub project, choose a predefined loop or describe your own, and the agent then writes its own notes about your product and suggests metrics to track. The author gave it a goal of growing contributions to a Telegram community. Loop broke that into KPIs and proposed an intake bot as one solution. That is one tester's report, so I would read it as a preview of the workflow and wait for more people to get in.
You can also reach Stitch Loop from Slack by mentioning @stitch for daily digests, or from coding agents like Antigravity through the Stitch CLI and MCP, according to Google.
Is this Jules under a new name?
TestingCatalog reports the internal name is Jitro, which it says was also the name of Jules v2. Google's post does not mention either one, so that link is unconfirmed. The timing is interesting all the same, because Google is shutting Jules down on December 31. If you still have Jules sessions, export them before then.
My read
I like the idea. Most software work happens after launch, and that is where coding agents have had the least to say. An agent that reads error logs and conversion numbers and comes back with a tested fix is aimed at the right part of the job.
I can also see how it goes wrong. An agent chasing a metric will find ways to move that metric, and some of them will be bad for customers. The review step only protects you if someone reads the pull request before clicking approve. And since this is a waitlisted Labs experiment with no price, nobody outside Google can say yet how well the suggestions hold up on a messy real codebase.
If you want something you can run today, Gemini 4 Carbon is already open to paying developers, and Claude Code Projects is Anthropic's version of an agent that keeps coding while you are away.
What this means for a small business
Stitch Loop is an agent that tells you what to fix next, where earlier ones waited to be told what to build. That only works when the goal is specific and the data is connected. A loop told to improve checkout conversion needs to see checkout data, and in a lot of small businesses that data is scattered across several tools.
Sorting that out is worth doing whether or not you ever use Stitch Loop. New Face Design offers a free process audit: we look at where your numbers live and which of your recurring fixes are worth handing to automation.