Google Stitch Loop turns live app metrics into GitHub pull requests

Google has introduced Stitch Loop, a Google Labs experiment that turns its Stitch design tool into a long-running product agent. Announced on November 18, 2026, Stitch Loop connects to a team's GitHub repository and to live product data such as logs, user behavior and feedback, then proposes improvements, writes the code in an isolated cloud sandbox and opens a GitHub pull request for a human to review. Google describes the project as a new direction for Stitch, which until now was known mainly as an AI tool for generating user interface designs.
The pitch addresses the stage of software work that coding agents usually leave alone. Most agents finish their job when a pull request is merged, while the maintenance that follows a release (security fixes, performance regressions, accessibility problems, conversion drops) keeps landing on small teams. Stitch Loop is built to keep watching a product after it ships, to measure whether its own changes moved the target metric, and to keep monitoring that metric after a goal has been reached in case it slips back.
Access starts with a waitlist that opened on November 18 for users worldwide, with priority for eligible Google AI Ultra subscribers. Google has not published pricing, usage limits or a general availability date, and every code change still requires approval, because nothing merges until the developer reviews the pull request.
··········
HOW STITCH LOOP WORKS: SUGGESTIONS, SOLUTIONS, BUILD AND REFLECT.
Google organizes the product around "loops", each tied to a goal, and every loop runs through the same four stages. The first stage surfaces suggestions: opportunities to move toward the goal, each backed by evidence drawn from the connected product data. The second stage turns a chosen suggestion into one or more possible solutions, with tradeoffs and scoping for each option and, where the change affects the interface, visual design alternatives produced with Stitch's design capabilities.
Once the developer picks a direction, the build stage begins. The agent writes the code in an isolated cloud sandbox and opens a pull request on GitHub, and Google says these pull requests can include screenshots or videos of the change when that helps the reviewer. The last stage, reflection, is the part that separates Stitch Loop from a conventional coding agent: the system monitors live data after the change is deployed, learns from the results and continues to watch the metric once the goal is met.
........
Stage | What Stitch Loop does | Human role |
Suggestions | Surfaces opportunities toward a goal, with evidence from product data | Chooses which suggestion to pursue |
Solutions | Proposes one or more options with tradeoffs, scoping and design variants | Picks a direction |
Build | Writes code in an isolated cloud sandbox and opens a GitHub pull request | Reviews and merges the pull request |
Reflect | Monitors live data, learns from results, keeps watching the metric | Decides whether to start a new cycle |
........
The approval step is the main safeguard Google describes. The agent never merges its own work, so a team keeps its existing review process, branch protections and continuous integration checks in front of every change Stitch Loop proposes. Google has not said whether the agent can be limited to specific directories or branches, or how it handles a pull request that a reviewer rejects.
··········
STITCH LOOP'S DEFAULT LOOPS AND THE GEMINI 4 CARBON SECURITY LOOP.
Stitch Loop ships with three default loops covering cybersecurity, latency and accessibility. These are areas where problems tend to accumulate quietly after launch and where improvement can be measured, through vulnerability findings, response times and accessibility checks. Teams can also define custom loops around their own goals, and Google's example is improving checkout conversion, a business metric that depends on analytics data rather than on the codebase alone.
The security loop is the only one for which Google names the underlying technology. It combines the cybersecurity capabilities of Gemini 4 Carbon with CodeMender, an autonomous code-security agent developed by Google DeepMind. According to Google, the loop flags vulnerabilities together with the evidence behind each finding and opens pull requests with proposed fixes. Gemini 4 Carbon is the model Google positioned for cybersecurity and long-horizon agent work when it launched it in October, a release covered in our report on Gemini 4 Carbon's launch for cyber defenders.
Google has not disclosed which models drive the latency, accessibility or custom loops, and it has not published benchmark results for Stitch Loop itself. Claims about how many vulnerabilities the security loop finds, or how much latency it removes, therefore cannot be checked yet.
··········
[[ADPLUS_300x250_1]]
STITCH LOOP INTEGRATIONS: GITHUB, 180+ DATA SOURCES, MCP, CLI AND SLACK.
Setting up Stitch Loop requires two kinds of connection. GitHub provides access to the codebase and is where the agent delivers its pull requests. Product context comes from logs, user behavior and feedback, which teams can connect through more than 180 integrations, through custom MCP servers, or by uploading data locally with a command-line interface. Google has not published the list of the 180-plus integrations in the announcement.
The product is also designed to feed other coding agents. Google says Stitch Loop can act as a service layer for agents such as Google Antigravity, and that any coding agent can pull the persistent product context Stitch Loop builds up through the Stitch CLI and MCP. That arrangement allows mixed setups in which Stitch Loop supplies the evidence about what users experience and a separate agent, possibly running a model from another provider, writes code inside the developer's usual environment.
Slack is the third surface. Mentioning @stitch in a Slack workspace returns daily digests of insights from the connected product, and teams can create new loops from Slack without opening the Stitch interface.
........
Connection | Purpose | Method |
GitHub | Codebase access and pull request delivery | GitHub integration |
Product data | Logs, user behavior, feedback | 180+ integrations, custom MCP servers, local CLI uploads |
Coding agents | Share persistent product context | Stitch CLI and MCP (e.g. Google Antigravity) |
Slack | Daily insight digests, new loops | @stitch mention |
........
··········
STITCH LOOP AVAILABILITY, WAITLIST AND AI ULTRA PRIORITY ACCESS.
Stitch Loop is available only through a waitlist at stitch.google.com/loop/waitlist, which opened on November 18 for all users globally. Eligible Google AI Ultra subscribers receive priority access, so Ultra members are likely to be among the first admitted, although Google has not said how many people it will let in or on what schedule. Feedback goes through the Google Labs Discord server, the channel Google uses for its Labs experiments.
The announcement does not include a price, a usage limit, a quota of loops or pull requests, or a date for broader availability. Google also does not say whether access will eventually require a paid Google AI plan or whether the free tier of Stitch will include it. Readers weighing an AI Ultra subscription partly for early access can compare what each plan includes in our Google AI Pro vs Google AI Ultra comparison.
........
Item | Status as of November 19, 2026 |
Waitlist | Open worldwide since November 18 |
Priority access | Eligible Google AI Ultra subscribers |
Price | Not disclosed |
Usage limits | Not disclosed |
General availability | No date announced |
Code changes | Delivered as GitHub pull requests, merged only after human review |
Required accounts | GitHub repository plus product data sources |
........
Teams that want to be ready when they are admitted can prepare in practical ways. The agent works through GitHub, so the repository needs the review rules and automated tests a team would apply to any outside contributor. The value of the reflection stage depends on product data, so teams should decide which logs, analytics and feedback sources they are willing to connect and check whether those tools appear among the supported integrations or can be exposed through an MCP server. Companies with security or compliance rules will also want to know where the cloud sandbox runs and how connected data is stored, details Google has not yet published. Background on the original design tool is in our guide to what Google Stitch is and how it works.
··········
FROM STITCH DESIGN TOOL TO STITCH LOOP: HOW THE PRODUCT HAS CHANGED SINCE I/O 2026.
At Google I/O in May 2026, Google presented Stitch as a design partner. The Stitch Agent streamed its work to the canvas in real time, users could steer iterations before a design was finished, inputs could be text prompts, speech, or existing code and design files, and finished designs could be exported to Google Antigravity for backend logic or published to the web through Netlify. Stitch Loop moves the product from producing designs before a launch to maintaining software after it.
........
Feature | Stitch (I/O update, May 2026) | Stitch Loop (November 2026) |
Main job | Generate and refine UI designs | Improve shipped products continuously |
Inputs | Prompts, voice, code and design files | GitHub, logs, user behavior, feedback |
Output | Designs, exports to Antigravity, Netlify publishing | GitHub pull requests with screenshots or videos |
After launch | No monitoring | Monitors live metrics and reacts to regressions |
Access | Available to global users | Waitlist, AI Ultra priority |
........
The timing also fits a reshuffle of Google's agent lineup. Google is retiring its Jules coding agent, which became read-only after November 11 and shuts down on December 31, as detailed in our report on the Jules shutdown, and is steering developers toward Google Antigravity. Stitch Loop takes a different slot in that lineup, working on what happens to code in production and passing its findings to coding agents such as Antigravity through the CLI and MCP.
··········
GOOGLE STITCH LOOP: CONFIRMED FACTS AND OPEN QUESTIONS.
Google has confirmed the core design of Stitch Loop: a four-stage cycle of suggestions, solutions, build and reflection; default loops for cybersecurity, latency and accessibility plus custom loops; a security loop that pairs Gemini 4 Carbon with Google DeepMind's CodeMender; connections to GitHub, more than 180 integrations, custom MCP servers, local CLI uploads and Slack; and a worldwide waitlist with priority for eligible AI Ultra subscribers. Every code change arrives as a GitHub pull request that a person must review before it merges.
The commercial and operational details are still missing. Google has given no price, no usage limits, no list of the 180-plus integrations, no model names for the non-security loops, no independent or internal performance figures, and no information about data handling inside the cloud sandbox. Until those details appear, Stitch Loop is a Labs experiment whose most concrete promise is a reviewed pull request generated from live product evidence, with the reflection stage still to be judged on real deployments.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org




