Most AI coding tools clock out the moment you push to production. Google’s Stitch Loop does not. Announced yesterday by Google Labs, Stitch Loop is a long-running product agent that connects simultaneously to your codebase and your live production metrics, identifies what needs fixing, writes the patch in an isolated cloud sandbox, and opens a GitHub pull request — then monitors your deployment to verify the fix actually worked. Nothing merges without human review. It runs continuously.
The framing matters: every major AI coding tool on the market has been competing to help you write code faster. Stitch Loop is competing to help you maintain what you already shipped. That is a different problem, and until now, nobody had built a continuous agent for it.
Three Loops, Three Gaps Most Teams Ignore
Stitch Loop ships with three built-in loops that address the three things engineering teams consistently defer: security, performance, and accessibility.
The security loop pairs Gemini 4 Carbon — Google’s cybersecurity-focused model — with CodeMender, an autonomous patching agent from Google DeepMind. CodeMender has already upstreamed 72 security fixes to open-source projects in its first six months; inside Stitch Loop, it scans your codebase proactively for vulnerability classes, not just known CVEs. It generates patches before incidents happen, not after.
The latency loop watches your live response-time data, identifies slow paths, opens targeted PRs, and then checks your production metrics after the deploy to confirm the improvement is real — not just a local benchmark win.
The accessibility loop monitors for WCAG compliance issues and opens PRs with screenshots and videos demonstrating what the fix looks like. For the majority of dev teams without a dedicated accessibility engineer, this one alone makes the waitlist worth joining.
How It Connects to Your Stack
You start by connecting GitHub and your product context — logs, user behavior data, analytics, feedback. Stitch Loop supports over 180 integrations for this, plus custom MCP servers and a local CLI for uploads. In Slack, the @stitch bot delivers daily insight digests and lets you kick off new loops directly from a message thread.
The MCP support is worth noting specifically. Google is positioning Stitch Loop as a service layer that other coding agents — including Google Antigravity — can pull context from. If you are already running a mixed-agent workflow, Stitch Loop can act as the persistent product-context backend for all of them.
The PR Approval Gate Is a Feature, Not a Limitation
Developer sentiment toward autonomous code-modifying agents runs about 40% skeptical, and the skepticism is not irrational. The concern is straightforward: an agent that opens a continuous stream of PRs without reliable signal about which ones actually matter shifts the burden from the people writing code to the people reviewing it. You have not automated maintenance — you have automated the creation of a review queue.
Google’s design choice — mandatory human approval before any merge — is a direct response to that concern. Code runs in an isolated cloud sandbox. Nothing touches your production environment until you say so. The agent proposes; your team disposes. This is the right architecture for a product still in experimental waitlist status, and frankly the right default for autonomous code changes at any maturity level.
What teams need to honestly assess before joining the waitlist: do you have the review bandwidth to handle a continuous stream of agent-generated PRs across three loops? If your PR review process is already a bottleneck, adding Stitch Loop will surface that problem fast.
Availability
The global waitlist is open today at stitch.google.com/loop/waitlist. Google AI Ultra subscribers get priority access. No public general availability date has been announced.
The three default loops cover the most common gaps. Custom loops — targeting things like test coverage, cloud cost, or deployment frequency — are the obvious next step, though Google has not confirmed a roadmap for them yet. Early developer reaction is mixed, with the security and accessibility loops drawing the most enthusiasm and the PR review burden drawing the most concern.

