Google Labs on Wednesday introduced Stitch Loop, an experimental agent that connects to a team's codebase and live product metrics and keeps proposing and building fixes after software ships, according to a company blog post. Google describes it as a long-running product agent that "keeps working 24/7 on your product until your goals are met."
Stitch Loop writes code in an isolated cloud sandbox and submits changes as GitHub pull requests. "Nothing merges until you review the PR," Justin Toupin, a director of product management at Google Labs, wrote in the announcement.
The release is an early example of coding agents moving from writing a feature to owning what happens to it in production. Google frames the gap directly. "Coding agents let you build and ship new features fast, but their job ends the moment your pull request merges," Toupin wrote.
How the loops work
Teams start by connecting GitHub and what Google calls product context, such as logs and user feedback. Google says this works through more than 180 integrations or custom MCP servers, and teams can also upload data from a command-line tool.
The agent then runs what Google calls evergreen loops, each tied to a measurable goal. According to the post, every loop moves through four steps.
Stitch Loop first suggests opportunities grounded in evidence from the product. It then offers one or more solutions with their tradeoffs, plus visual design options where they apply.
Once a person picks a direction, the agent builds the change in the sandbox and opens a pull request, which can include screenshots and videos of the work. In the last step it monitors live data and learns from the results. Once a goal is met, it keeps watching in case the metric slips, Google says.
Built-in security loop
Stitch Loop ships with default loops for cybersecurity and latency, plus one for accessibility. Users can add custom loops, with improving checkout conversion given as an example. Google singled out security as "often the hardest gap to cover for lean teams shipping quickly."
The security loop pairs Gemini 4 Carbon with CodeMender, which Google describes as Google DeepMind's autonomous agent for improving code security. Google says the pair scans a codebase for vulnerabilities, flags each issue with evidence and opens a pull request with a proposed fix for review.
A layer for other agents
Google is also pitching Stitch Loop as a service that other coding agents can call. Agents such as Google Antigravity can pull persistent product context from Stitch Loop through the Stitch CLI and MCP, which Google says supports mixed-model setups. In Slack, teams can mention @stitch to get daily insight digests and create new loops.
The tool extends Stitch, which Google says millions of people already use to turn ideas into working interfaces. Google calls Stitch and Stitch Loop steps toward agents that help lean teams from ideation through continuous optimization.
Access and open questions
Stitch Loop is not generally available. Google opened a global waitlist on Wednesday, with priority for eligible Google AI Ultra subscribers. Josh Woodward of Google Labs wrote on X that the team has "iterated through a few versions" and that "it's still early, so we're going waitlist on this one."
TestingCatalog, which said it had tested Loop for weeks, reported that its early access ran on Gemini 3.8 Flash. Google's announcement did not give pricing, say how often loops run by default, or publish accuracy figures for the fixes the agent proposes. Human approval of each pull request is the only control the post describes in detail.
Source: Google, “Introducing Stitch Loop: A new way to build and evolve products”
Additional sources: Stitch by Google on X, “Introducing Stitch Loop”; TestingCatalog, “Early look at Loop, new autonomous AI agent from Google”
