Google Stitch Loop: The Product Agent That Keeps Improving Your App After Launch
Google Labs extends Stitch into post-launch optimization with security, latency, and accessibility loops

Google Labs on November 18 introduced Stitch Loop, an always-on product agent that connects to a live codebase and production metrics, identifies problems in security, latency, and accessibility, and opens GitHub pull requests with proposed fixes—automatically, around the clock, until the team decides to stop it. Where today's coding agents sign off the moment a pull request merges, Stitch Loop treats launch as the beginning of the job rather than the end.
The announcement extends Google's Stitch platform, an AI-native design and prototyping tool that launched at Google I/O 2025 and was substantially rebuilt in March 2026 as a "vibe design" canvas. The product is in experimental access, with a global waitlist open at stitch.google.com/loop/waitlist and priority access for eligible Google AI Ultra subscribers.
The Boundary Every Coding Agent Stops At
GitHub's Copilot Coding Agent, which reached general availability in September 2025, works asynchronously on assigned issues and delivers a draft pull request—then stops. Cognition's Devin, Cursor, Claude Code, and Google's own Antigravity development platform all share this structure: a task goes in, a PR comes out, and the agent moves on.
What none of these tools do is monitor what happens after the PR merges—whether a vulnerability surfaced, whether a recent deploy introduced latency regressions, whether the checkout conversion rate dropped. That work falls to observability platforms, security scanners, and human engineering triage. Stitch Loop's claim is that an AI agent can own that work instead, completing the loop from live production signal to proposed code change without waiting for a ticket to be filed.
How the Four-Step Cycle Works
Each loop inside Stitch Loop runs the same four-stage cycle. First, the agent surfaces suggestions: opportunities aligned with the loop's goal, backed by evidence from connected product data. Each suggestion becomes one or more solutions, presented with tradeoffs and—where the loop involves UI changes—visual design options. The agent then enters the build phase: it writes the proposed code in an isolated cloud sandbox and opens a GitHub pull request, including screenshots or short videos where relevant. After a PR is reviewed and merged, the agent reflects—monitoring live metrics against the goal and continuing to watch even after the goal is met, in case a later change causes regression.
Nothing auto-merges. Google is explicit that the team has final approval on every pull request before any change reaches production.
Teams connect their codebase through GitHub and feed product context through one of more than 180 integrations, through custom MCP servers, or through local uploads via the Stitch command-line interface. A Slack integration delivers daily insight digests and allows new loops to be created from within the platform.
Gemini 4 Carbon and CodeMender Power the Security Loop
Three default loops ship with Stitch Loop: cybersecurity, page-load latency, and WCAG accessibility compliance. The security loop is built on a combination of Gemini 4 Carbon and CodeMender.
Gemini 4 Carbon was announced on October 28, 2026 as Google's current frontier model, with reported state-of-the-art performance across cybersecurity defense—vulnerability detection, validation, and remediation. In company-run tests, Carbon achieved more than twice the secure-code generation success rate of its predecessor, Gemini 4 Argon. Access at launch was restricted to cybersecurity defenders through Google's Fairwind Program; Stitch Loop is among the first developer-facing products to deploy it in a production agent workflow.
CodeMender was developed by Google DeepMind as an autonomous vulnerability-remediation agent. It automatically patches new vulnerabilities as they appear and rewrites existing code to remove entire classes of vulnerabilities. The agent uses advanced program analysis and multi-agent systems, confirming that vulnerabilities are exploitable before proposing fixes—only surfacing patches that pass its validation for human review. A built-in LLM-based critique agent compares the original and modified code before a fix is proposed. Before Google moved it to enterprise preview on the Gemini Enterprise Agent Platform, CodeMender had contributed 72 security fixes to open-source projects, including some as large as 4.5 million lines of code.
Inside Stitch Loop's security loop, Carbon's vulnerability-detection reasoning and CodeMender's remediation pipeline work together—identifying a problem, confirming it is exploitable, writing a fix, and packaging it as a pull request with supporting evidence. The models powering the latency and accessibility loops have not been disclosed.
Stitch Loop as a Context Layer for Other Agents
A detail that separates Stitch Loop architecturally from a simple scheduled-scan tool is what it accumulates: a persistent, growing record of product context—codebase understanding, metric history, improvement record—that other agents can access. Through the Stitch CLI and MCP server, external coding agents can pull this context directly. Google cites Antigravity as an agent that can use Stitch Loop as a service layer when executing discrete development tasks.
The result is a potential division of labor across Google's agentic developer stack: Stitch generates interfaces, Antigravity implements and runs code, and Stitch Loop maintains a live model of the shipped product and proposes continuous improvements. Each layer exposes its context to the others through MCP, turning what were separate tools into a loosely coupled multi-agent pipeline.
Limitations and What Remains Unproven
Stitch Loop carries significant caveats. It is a Google Labs experiment with no pricing announced and no general availability timeline. The accuracy of loop outputs—false-positive rates on security findings, the precision of latency suggestions, the appropriateness of accessibility change proposals—has not been independently benchmarked. An early hands-on reviewer noted that the agent produced unsolicited theme changes during testing, a reminder that loosely scoped loop goals can generate unexpected side effects and that human PR review is not a formality.
The move from an isolated cloud sandbox to a merged production change involves a hand-off that Google has not described in detail. The data handling specifics for the Stitch Loop sandbox—including whether source code is retained, as it is not in the enterprise version of CodeMender—have not been published for this product.
Whether Google can sustain the operational reliability that an always-on product agent requires—consistent pull-request quality, interpretable reasoning for proposed changes, and stable long-horizon execution—remains to be demonstrated at scale. The teams most likely to benefit first are those already on GitHub, already using Stitch, and facing post-launch security or performance debt that currently has no automated owner.



