Introducing Stitch Loop: A new way to build and evolve products

Open original

Google Labs is launching Stitch Loop, an agent that connects to your codebase and live metrics, finds problems after code ships, and proposes fixes as GitHub pull requests for your approval. It runs on loops for security, latency, accessibility, and custom goals like checkout conversion.

  • Stitch Loop runs four steps per loop: surface suggestions from product data, offer solutions with tradeoffs, build the fix in a cloud sandbox and open a PR, then monitor live metrics to see if the goal holds.
  • The security loop combines Gemini 4 Carbon's cybersecurity capabilities with CodeMender, DeepMind's autonomous code-security agent, and flags each issue with evidence and a proposed fix.
  • It plugs into existing tools: coding agents like Google Antigravity can pull its product context through the Stitch CLI and MCP, and teams can get daily insight digests in Slack. The waitlist opened globally, with Google AI Ultra users getting priority.

Stitch Loop is a Google Labs product agent that picks up where coding agents stop. Coding agents can build and ship features quickly, but their work ends when a pull request merges. Stitch Loop connects to your GitHub repo and your live product context, which can include logs, user behavior data, and feedback, through more than 180 integrations, custom MCP servers, or local CLI uploads. It then runs continuous loops aimed at measurable goals. Three default loops come built in, covering cybersecurity, latency, and accessibility, and you can add custom ones, such as improving checkout conversion or fixing signup drop-off. Each loop follows the same four steps. It surfaces suggestions backed by evidence from your product, offers one or more solutions with tradeoffs and scoping (plus visual design options where relevant), builds the chosen fix in an isolated cloud sandbox and opens a GitHub pull request complete with screenshots or videos of the work, and then monitors live data to see whether the change moved the metric. Once a goal is hit, it keeps watching in case the number slips back. Nothing merges without your approval. The security loop is the most concrete piece: it pairs Gemini 4 Carbon's cybersecurity capabilities with CodeMender, a Google DeepMind agent for code security, to scan for vulnerabilities and propose fixes as PRs, each flagged with supporting evidence. Access and integration are the other practical details. Stitch Loop can serve as a context layer for coding agents like Google Antigravity, feeding them persistent product context through the Stitch CLI and MCP so teams can mix models or automate improvements inside existing workflows. It also runs in Slack, where @stitch delivers daily insight digests and lets you create new loops. The waitlist opened today for all users globally, with eligible Google AI Ultra subscribers getting priority access. Sign-up is at stitch.google.com/loop/waitlist, and feedback goes to the Google Labs Discord. The article offers no pricing, performance benchmarks, or data on how the loops perform in practice, so the claims about results remain untested until people with real codebases start using it.

Ask about this article