Google Stitch Loop: An Agent That Stays After Launch

Most coding agents clock out the moment your pull request merges. Google's new Stitch Loop is built to do the opposite: it hangs around after launch, watches your live numbers, and keeps sending fixes. Google Labs announced it on November 18, 2026, and I think the idea matters more than the product.
What Stitch Loop actually is
Google calls Stitch Loop "a long-running product agent." In plain words: you hook it up to your GitHub repo and to whatever tells you how your product is doing, such as logs, user behaviour and feedback. Google says it plugs into more than 180 integrations, plus custom MCP servers or files you upload from a local CLI.
Then you give it a goal you can measure. Faster pages. Fewer security holes. More people finishing checkout. It works toward that goal in a loop with four steps:
- Suggestions. It points at things worth fixing, backed by evidence from your product.
- Solutions. Each suggestion comes with one or more fixes, with tradeoffs and scope spelled out.
- Build. You pick one. It writes the code in an isolated cloud sandbox and opens a GitHub pull request, with screenshots or videos when that helps.
- Reflect. It watches the live data to see whether the change worked, learns from the result, and keeps watching after the goal is met in case the number slips.
The part I like most is the boring rule at the end: nothing merges until a human reviews the PR.
A quick example
Say you run a small store and checkout drop-off has crept up. You set a custom loop with the goal "raise checkout conversion." Stitch Loop digs through your analytics, spots that mobile users bail on the address form, and offers two options: cut the form down, or add address autocomplete. You pick autocomplete. A PR lands with screenshots. You merge it. Then the loop keeps an eye on the conversion number, and if it goes back down three weeks later, it comes back with fresh suggestions.
That last step is the whole pitch. Today the "did it work?" question usually lives in a dashboard nobody opens on Fridays.
Three loops out of the box
You get three default loops: security, latency and accessibility. Coverage from TestingCatalog says the accessibility loop targets WCAG. The security loop pairs Gemini 4 Carbon with CodeMender, the Google DeepMind agent built to find and patch code vulnerabilities. Each security issue comes flagged with evidence and a proposed fix as a PR.
It also works as a service layer for other coding agents. Google says any agent you like can pull product context from it through the Stitch CLI and MCP, and there is an @stitch bot in Slack for daily digests and for starting new loops.
My take
The useful shift here is in the goal. Most coding agents get a task: "add this button." Stitch Loop gets an outcome: "make checkout better." That's closer to real product work, and harder to fake. A metric either moves or it doesn't.
But I'd keep two worries in mind. First, an agent that optimises a single number will happily optimise the wrong one. Point it at "time on page" and you may get a slower, more confusing page. Pick your goals the way you'd pick them for a new hire who takes everything literally. Second, "a human approves every PR" only helps if that human reads the PR. A steady stream of small, reasonable-looking fixes from a tireless agent is exactly the kind of thing people stop reading closely around week three.
Right now it's an experiment with a waitlist, open globally, and eligible Google AI Ultra users get priority access. So there are no real-world results to judge yet.
Still, the direction feels right to me. Shipping was never the finish line. If agents start owning the "after" part, the job of the developer looks a lot more like choosing goals and saying no.
So here's my question: which number in your product would you trust an agent to chase on its own, and which one would you never hand over?