Google Labs launched the experimental product Stitch Loop. This AI agent continues to monitor online numbers, find problems, write corrections, and open PRs for approval after the product is launched. It has three built-in tasks: information security, speed, and accessibility, and the waiting list is open globally. (Previous summary: Google’s most powerful model Gemini 4 Carbon is online! Specializing in long-term reasoning, the security vulnerability detection rate increased by 25%) (Background supplement: Google Stitch open source design system: DESIGN.md allows Claude Code, Codex, and Antigravity to generate high-quality UI) After the product is online, the AI ​​programming tool usually stops working. However, the Stitch Loop (Stitch is Google's AI interface design tool) launched by Google Labs, Google's experimental product department, is different: after it goes online, it continues to stare at the online numbers, identify problems, write corrections, open PR, and wait for approval until it reaches the goals set by the user. The official Stitch account was made public on X on November 18, and the official blog also stated it on the same day. The official statement is: "Stitch helps you design your product, and Stitch Loop helps you continue to improve it after it goes online." Introducing Stitch Loop, a new Stitch experiment. Stitch helps you design your product. by Google (@stitchbygoogle) November 18, 2026 How to actually do it: Before using it, you must first connect to GitHub, and then connect product data such as logs, user behavior, feedback, etc. There are more than 180 ready-made connections, custom MCP servers (standard interfaces for AI tools to connect to external data), or upload from the local machine. Stitch Loop has 3 built-in preset tasks, officially called loops, which refer to improvement tasks that are repeatedly executed until the target is reached. Users can also add new ones by themselves. The process of each task is as follows: it first points out which areas can advance the goal based on the evidence in the product; then attaches one or more solutions to each suggestion, stating the trade-offs and scope of work; after the user chooses the direction, it writes the program in an isolated cloud sandbox (a test space separate from the formal environment), then opens a PR on GitHub and waits for approval, and attaches screenshots and videos when relevant; after the merger is launched, it continues to look at the online numbers, learns from successes and failures, and will keep an eye on it after the goal is achieved to prevent the indicator from falling back. Each of the three built-in tasks manages one. The security task uses the security capabilities of Gemini 4 Carbon, Google's current strongest model, plus Google DeepMind's CodeMender, which is a security patching agent released in October 2025. In the first six months, 72 patches have been adopted by open source projects. The speed task identifies slowdowns from online data, and the accessibility task focuses on WCAG (international web accessibility standards that allow users with visual impairments and other disabilities to use them) issues. Will they randomly modify the official product? The official emphasizes that the final decision lies with the user. PR will not enter the official version without being reviewed by others. The example in the promotional video shows how it works: the user sets the goal to "find and fix every known vulnerability before the vulnerability affects the user." Loop finds the /api/orders endpoint and will return all customers' orders without logging in, and proposes a correction: only return the orders of logged-in customers. Who can use it, and what else is unknown? The waiting list will be open to the world from November 18, and eligible subscribers of Google AI Ultra (Google’s highest-end AI subscription plan) will be given priority. New startups in the seed round or Series A that have already launched products can apply to be "design partners" and work directly with the development team. It's still experimental. The official said that this idea has been revised several times and is "still in the early stages". It will be opened to more people if it goes well. The price is not disclosed, and there are no performance figures such as how many bugs have been fixed or how many indicators have been improved. Moreover, the program code and user data must be connected before it can operate. The only external trial experience so far comes from TestingCatalog: The author tried it for several weeks before launch. At that time, Loop was running the old model Gemini 3.8 Flash. As a result, he made many small adjustments to the layout of his own website and made suggestions on accessibility and compliance. Related reports Google launches Antigravity 2.0: Gemini 3.5 Flash engine, dynamic sub-agent, and Scheduled scheduling functions are in place Google launches Gemini 4 Hydrogen! 4 levels of thinking depth, focusing on daily programming and AI agents Claude Design overhaul: importing the design system, opening up the code in both directions, alleviating the nightmare of eating tokens Google’s promotion system has been changed to “AI performance”, and employees have broken the product ecosystem into pieces Google launches Gemini 4 Neon benchmark model, which can be used with Pro subscription〈Google’s new Stitch Loop: After the product is launched, it continues to find loopholes and fix problems> This article was first published in Dongqu BlockTempo "Dongqu Dongzhu - the most influential blockchain news media".