November 19, 2026AgentsCodingTool

Stitch Loop: Google Wants the Agent to Stay After the Merge

Coding agents have a weird job description: they work hard until the pull request merges, then forget the product exists. Google Labs' answer, announced Nov 18, is Stitch Loop, an experimental long-running product agent that connects to your codebase and live metrics and keeps working toward a goal. Announcement: https://blog.google/innovation-and-ai/models-and-research/google-labs/stitch-loop/

The setup: connect GitHub plus product context, meaning logs, user behavior and feedback, through 180+ integrations, custom MCP servers or local CLI uploads. Then Loop runs what Google calls evergreen loops. Three come built in (security, latency, accessibility) and you can add your own, like checkout conversion. Each loop goes through four steps. Suggestions grounded in evidence from your product. Solutions with tradeoffs, scoping and design options. Build, where it writes code in an isolated cloud sandbox and opens a GitHub PR, with screenshots or video when relevant. Reflect, where it watches live data, learns from wins and misses, and keeps watching after the goal is hit in case the metric slips.

The guardrail is the old one: nothing merges until a human reviews the PR. The security loop pairs Gemini 4 Carbon with DeepMind's CodeMender. Other agents can pull Loop's product context through the Stitch CLI and MCP, so it can sit as a service layer under Antigravity or anyone else's coding agent, and @stitch in Slack sends daily digests.

The interesting shift is who owns the goal. Until now the human owned the metric and the agent owned the ticket. Loop flips that: the agent holds the metric and proposes the tickets, the human holds only the merge button. That is a lot of trust to put in "evidence grounded" suggestions, and Google has published no numbers on how often its loops actually move a metric. It is a waitlist experiment for now (https://stitch.google.com/loop/waitlist), with Google AI Ultra users first. Worth watching less for the product than for the pattern: the agent's job no longer ends at merge, so its mistakes don't either.
← Previous
Three Papers on the Agent That Says Done and Keeps Going
Next β†’
KPMG Put 100 Agents on a Real Audit. They Asked Better Questions
← Back to all articles

Comments

Loading...
>_