Google Stitch Loop: always-on product agent

Google Labs introduced Stitch Loop on November 18, 2026 - a long-running product agent that connects to a codebase and live product metrics, looks for concrete improvements, writes code in an isolated cloud sandbox and opens a GitHub pull request for human approval. Access is via a global waitlist, with Google AI Ultra subscribers getting priority.
- A product agent, not a chat bot — Stitch Loop connects to a GitHub repository plus product context such as logs, user behaviour and feedback, then runs continuous 'evergreen loops' against measurable goals instead of waiting for a prompt.
- Four-step loop: suggest, solve, build, reflect — Each loop proposes opportunities, offers solutions with trade-offs, builds the change in an isolated cloud sandbox and opens a GitHub pull request, then monitors live data and keeps watching for regressions after the goal is met.
- Three default loops — Google ships built-in loops for cybersecurity, latency and accessibility, and lets teams add custom loops such as improving checkout conversion.
- Human approval is the merge gate — Google is explicit that nothing merges until a person reviews the pull request, and that the agent's code is written in a sandbox rather than directly on the main branch.
How the loop actually works
Related guide: Microsoft Ignite 2026: Agent 365's observability turn
The design is best understood as a repeating four-step cycle rather than a single agent run. Google describes the first step as Suggestions: the agent surfaces an opportunity to make progress toward a goal, grounded in evidence from the product rather than in a generic idea of what might be good. The second step is Solutions, where each suggestion is paired with one or more proposed solutions that include their trade-offs and scoping, plus visual design options where those are relevant.
The third step is Build. Once a direction is chosen, Stitch Loop writes the code in an isolated cloud sandbox and opens a GitHub pull request. Where relevant, that pull request can include screenshots and videos of the change. The final step is Reflect: the agent monitors live product data as it works toward the goal, using positive and negative results to sharpen later rounds of suggestions. Once the goal is reached, it does not shut down - it keeps watching in case the metric slips again.
That last detail is the real product idea. Most coding agents are one-shot tools whose value is consumed at the moment of execution. Stitch Loop is designed around the opposite assumption: that the interesting work in a product happens after launch, as a long tail of small regressions, security patches and performance improvements that rarely justify a full sprint each.
What it connects to, and what it will not touch
Related guide: Microsoft Execution Containers: Agent Security, Explained
The agent needs context to work with, and Google describes three ways to supply it. Teams can connect GitHub and their product context - logs, user behaviour, feedback and similar signals - through more than 180 integrations. They can add custom MCP servers for systems the integration list does not cover, or they can upload context locally through a CLI. That last option matters for teams that cannot send logs to a third-party service.
Google says Stitch Loop can be used as a service layer for other coding agents, including Google Antigravity, and that other agents can pull persistent product context from Stitch Loop through the Stitch CLI and MCP, which is how mixed-model setups are meant to work. Inside Slack, the @stitch mention is used for daily insight digests and for creating new loops - a sign that Google expects this to become part of a team's routine rather than a separate console nobody opens.
The governance story is the important part for anyone considering it. Google's own summary of the launch states it plainly: the tool identifies issues, writes code in a sandbox, and submits pull requests, and the team keeps full control because every pull request must be approved manually. Nothing merges on its own. That is a meaningful distinction from fully autonomous maintenance agents, and it also means the review workload moves rather than disappears - the agent proposes a larger number of small, already-implemented changes for a human to accept or reject.
The security loop is the flagship
Related guide: Gemini 4 Carbon: Google's Cybersecurity-First Frontier at 88.2% SEC-bench
Three default loops ship with the product: one for cybersecurity, one for latency and one for accessibility. Teams can add their own at any time - Google gives improving checkout conversion as the example - and the target is always a measurable outcome rather than a vague instruction to make things better.
The security loop is the most concrete, because Google says it pairs Gemini 4 Carbon's cybersecurity capabilities with CodeMender, Google DeepMind's autonomous agent for improving code security. The loop scans the codebase for vulnerabilities before they become incidents, flags each issue with supporting evidence, and opens a pull request with a proposed fix for review. The pattern matters: it is not a scanner that ends in a report you have to triage, but a scanner with a remediation path already attached.
For a lean team, that is the argument for the whole product. If the security loop is reliable, the hard part of dependency scanning is not finding the advisory, it is turning a CVE into a reviewed patch while the rest of the roadmap keeps moving. Stitch Loop's contribution is the plumbing between detection, context, code change and review.
What to check before you trust it with a repository
Related guide: Gemini 4 Hydrogen: Google's faster agent model explained
Stitch Loop is an experiment with a waitlist, so there are practical questions a team should answer before connecting a production repository. The first is scope of write access: the agent can write code and open pull requests, so the repository permissions, branch protection rules and required-reviewer configuration are the real safety boundary, not the approval step alone. The second is data exposure: the product context includes logs and user-behaviour data, and teams should establish which of that context can travel to a cloud sandbox before connecting production telemetry.
The third is goal design. A loop that optimises a metric can optimise it in ways a human reviewer would not accept. Google's own framing - suggestions grounded in evidence, solutions with trade-offs, human final call - is the correct shape, but the effectiveness depends entirely on whether the goal and the guardrails are written down before the agent starts.
Finally, teams should decide in advance how the loop's pull requests enter the normal review workflow. If they arrive in a separate queue that nobody owns until the queue becomes annoying, the design fails quietly. The most practical setup is likely to require a human owner for each loop, a defined threshold for when a suggestion is worth reviewing, and a rule that the agent's own changes are never self-approved.
Where this fits against the rest of the agent stack
Related guide: Best VPS for AI agents in 2026: Hetzner vs AWS Lightsail vs EC2
Stitch Loop is not competing with coding agents for feature work; it is aimed at the unglamorous 90 percent of a product's life after the launch post. That makes it closer to a continuous performance and security engineer than to an assistant, and it explains why the product starts with evidence-grounded suggestions rather than a chat box. It also explains the emphasis on approval: the value of the product comes from volume of small changes, which only works if each change is cheap to verify.
For individual developers, the waitlist is worth joining if you already have a real repository with real telemetry, because that is the input the product is designed around. For teams, the interesting part is the workflow design: the agent is the easy part, and deciding which loops to enable, what data they may see and who reviews their output is the work that determines whether it helps or becomes noise. Teams that already run their own workflow automation and monitoring will find the integration question familiar, because the hard part of any always-on agent is the same: a measurable goal, a bounded context and a human who owns the outcome.
Frequently asked questions
A product agent, not a chat bot
Stitch Loop connects to a GitHub repository plus product context such as logs, user behaviour and feedback, then runs continuous 'evergreen loops' against measurable goals instead of waiting for a prompt.
Four-step loop: suggest, solve, build, reflect
Each loop proposes opportunities, offers solutions with trade-offs, builds the change in an isolated cloud sandbox and opens a GitHub pull request, then monitors live data and keeps watching for regressions after the goal is met.
Three default loops
Google ships built-in loops for cybersecurity, latency and accessibility, and lets teams add custom loops such as improving checkout conversion.
Human approval is the merge gate
Google is explicit that nothing merges until a person reviews the pull request, and that the agent's code is written in a sandbox rather than directly on the main branch.
Prices and benchmark figures are standard published rates as of November 19, 2026 and can change without notice. Some outbound links may be affiliate links — see our disclosure. Found an error? Tell us and we'll correct it in public.