On November 18, 2026, Google Labs launched Stitch Loop, an experimental product and an AI agent that 'keeps working after launch.' It connects to a code repository and live product data, keeps finding issues and writing fixes, and hands them over as GitHub pull requests. Coding agents are plentiful, but one that watches a shipped product and keeps iterating on real data is a first.
How it works
Stitch Loop works in four steps: suggest improvements grounded in product data, propose a solution with trade-offs, write the code in an isolated cloud sandbox, then check the result against live metrics after deployment. It offers more than 180 integrations, plus custom MCP servers and local CLI uploads. Three built-in loops cover security, latency and accessibility; the security loop pairs Gemini 4 Carbon with DeepMind's CodeMender to scan for code vulnerabilities. Every change must be reviewed by a human before it merges — a point Google repeats throughout the launch announcement.
Join the waitlist — not everyone gets in right away
For now, Stitch Loop is only an experiment. A global waitlist is open, with eligible Google AI Ultra subscribers getting priority access. No pricing and no date for wider availability have been announced, and details may change as testing continues. Google also released Gemini 4 Neon the same day; both launches point in the same direction: after the model release, the real competition is over who can keep maintaining the product.
For development teams, the appeal of Stitch Loop is not writing code, but finally having someone cover the vacuum after launch, when nobody is watching the shipped product. The security loop's value is the most direct: vulnerabilities get flagged and turned into fixes before they become incidents. Whether it can tell the difference between 'metrics got worse' and 'users just don't like it' will only become clear once waitlisted users run it for real. Before joining the queue, one question is worth settling: are you willing to let an agent read your codebase and user data on an ongoing basis? That costs far more trust than a one-off code generation.