Google introduces Stitch Loop to improve products
Google Labs has introduced Stitch Loop, an agent that connects code with a product's metrics and feedback to spot problems, suggest solutions, and open changes in GitHub. The waitlist is now open, and eligible Google AI Ultra users will get priority.

Google Labs has introduced Stitch Loop, an agent that connects to your code and product metrics to spot problems, suggest changes, and continuously open code requests. The waitlist is now open worldwide, although eligible Google AI Ultra users will get priority.
The idea starts with a problem familiar to every software team: building a feature is only the beginning. Then come security bugs, screens that load too slowly, and users who abandon a sign-up or purchase without explaining why.
Coding agents typically stop working once a code request is accepted. Stitch Loop aims to cover what happens next by continuously tracking the product and its results.
How Stitch Loop works
First, you connect GitHub and your product context. This can include technical logs, user behavior, and feedback through more than 180 integrations, custom MCP servers, or uploads from the local command line.
MCP is a standard that lets an agent connect to external tools and data sources. In this case, it keeps Stitch Loop from working blindly by giving it real information about your application.
Using that data, the agent works in continuous cycles. Each cycle follows four steps:
- Suggestions: identifies opportunities related to the goals you have defined, based on evidence from the product.
- Solutions: suggests one or more ways to address each problem, along with their benefits, costs, and scope. When relevant, it also includes visual design options.
- Build: once you choose a direction, it writes the code in an isolated cloud environment and creates a change request in GitHub. It can include screenshots and videos of the result.
- Review: monitors product data after the change and uses the results, positive or negative, to adjust the next round of suggestions.
The system initially includes cycles for security, loading speed, and accessibility. You can also create others, such as one focused on increasing the proportion of users who complete a purchase.
Nothing is added to the product automatically. The change request remains pending your review and approval, so you retain the final say over what goes into production.
A watchdog for security and performance
The security feature is designed to find vulnerabilities before they cause an incident. Stitch Loop combines the cybersecurity capabilities of Gemini 4 Carbon with CodeMender, an autonomous Google DeepMind agent focused on improving code security.
When it detects a problem, it shows the evidence supporting its finding and prepares a proposed fix in GitHub. The team decides whether to accept, modify, or discard it.
How it fits with your tools
Stitch Loop is not designed to force you to abandon the agents you already use. It can work as a service layer for tools such as Google Antigravity and share context through Stitch CLI and MCP with configurations that combine different models.
It also integrates with Slack. By mentioning @stitch, you can receive daily summaries about the product's status or create new work cycles from the same channel where your team already works.
For you, the potential change is concrete: instead of manually reviewing every metric, error, or comment to decide what to improve next, an agent can organize those signals, prepare alternatives, and turn the option you choose into reviewable code. That does not eliminate human oversight, but it can reduce repetitive work for small teams.
The waitlist is available globally at stitch.google.com/loop/waitlist. The important question now is how much access the agent gets to each product's data and how much value its cycles deliver once the changes reach real users.

