Google Labs has unveiled Stitch Loop, an AI agent that keeps improving software after it ships, and waitlist signups opened worldwide on Wednesday — this week, not someday.
The tool links a live product’s codebase and real-time metrics, then runs continuous fix cycles against goals the user sets, Google said in its announcement on November 18.
AI Matters reported the details of the launch, which Google positions as a new experiment inside Stitch, the Labs tool that turns an idea into a working screen. Stitch already counts designers, developers and founders in the millions, per the company’s own blog post. Stitch Loop is aimed at what happens after the merge.
That framing matters. Google’s argument is blunt: a coding agent’s job ends the moment a pull request lands, but a shipped product never stops needing work — vulnerability patches, faster checkout, figuring out why signups churn. Stitch Loop exists to own that post-launch grind, which today eats most engineering teams’ calendars.
The mechanism is a four-stage loop. Proposal: the agent surfaces improvement opportunities backed by product data. Solution: it presents one or more fixes, complete with trade-offs, scope, and design mockups where relevant.
Build: chosen solutions get written in an isolated cloud sandbox and packaged as GitHub pull requests for human approval. Retrospective: the agent watches live product data, feeds results into the next proposal, and keeps monitoring after a target is hit in case the metric degrades.
Three loops ship first
Out of the gate, there are three default loops — cybersecurity, latency, and accessibility — plus custom loops for anything else, such as improving payment conversion. Users connect product context through more than 180 integrations, custom MCP servers, or local CLI uploads: GitHub repos, logs, user behavior, feedback streams. The agent then optimizes against whatever metric the user designates.
The security loop is the one enterprise eyes should linger on. It pairs Gemini 4 Carbon’s cybersecurity capabilities with CodeMender, Google DeepMind’s autonomous agent for code security, to hunt vulnerabilities across the codebase. Every finding comes with cited evidence, and each fix ships as its own pull request.
One guardrail stands out in the announcement, and it’s the right one: no code merges until a human reviews the pull request. That’s the difference between an autonomous agent and an autonomous liability. In an era where AI-generated code is already flowing into production at scale, keeping the merge gate human is the design decision that determines whether tools like this get enterprise adoption or get blocked by security teams.
Where it plugs into the stack
Google is explicit that Stitch Loop can serve as a service layer on top of coding agents like Google Antigravity. Other agents aren’t locked out either — they can pull product information through Stitch CLI and MCP. That’s a quietly significant architectural choice.
Rather than building another walled coding agent, Google is building the telemetry-and-iteration layer that any agent can consume, which puts Stitch Loop in a position to become plumbing for the wider agent ecosystem, not just another Google product.
Slack integration rounds it out: invoke @stitch and you get a daily analysis summary, plus the ability to spin up new loops from chat. It’s a small feature, but it signals where this lives — inside the team’s existing workflow, not behind another dashboard login.
Access starts with Google AI Ultra subscribers who meet eligibility conditions, with a global waitlist open to everyone as of this week’s announcement.
What this changes for developers
The competitive context is hard to ignore. GitHub Copilot, Cursor, and a wave of agentic coding tools have been racing to automate the write-and-merge cycle.
Stitch Loop’s pitch is that the merge is the beginning, not the end — that the real backlog lives in production telemetry, not in the issue tracker. If it works as described, the bottleneck shifts from writing code to reviewing it, which makes human review capacity the scarce resource.
It also tightens Google’s grip on the full lifecycle. Stitch takes you from idea to screen; Stitch Loop takes you from screen to maintained product; Gemini 4 Carbon and CodeMender handle the security sweep underneath, according to Aimatters.
For teams already inside Google’s ecosystem, the gravity is obvious. For everyone else, the MCP and CLI doors are the escape hatches — and whether Google keeps those genuinely open is the thing to watch as the experiment matures.
Skeptics will note the usual caveats around Labs experiments: this is explicitly labeled an experiment, gated behind a waitlist, and available first only to top-tier subscribers. Google’s track record with Labs projects is mixed, and nothing here promises general availability. The company didn’t disclose pricing beyond the AI Ultra requirement, token costs for running continuous loops, or which integrations ship on day one versus later.
Still, the concept lands. Post-launch maintenance is the least glamorous, most expensive part of software — and it’s the first problem an agent can tackle with truly objective feedback, because production metrics don’t lie about whether a fix worked. The retrospective stage closing the loop back into the next proposal is the piece that turns this from a code generator into a genuine optimization system.
Stitch Loop’s home is a fitting one. Google Labs, the company’s experimental arm, dates back to 2002 and served as the proving ground for products that later became household names — Gmail and Google Calendar both passed through it.
The original Labs was shut down in 2011 as Google streamlined its product lineup, but the branding was revived in the 2020s, publicly relaunching in 2023 with a focus on generative AI experiments such as NotebookLM and MusicFX. Today it hosts experimental AI products including Veo, Search Labs, and Google Stitch itself.
Whether Stitch Loop earns a permanent place in the toolchain will show up in the only metric that matters: does the pull request queue fill with fixes that real users can feel? The waitlist is open, and the answer starts arriving once the first merge happens.