Handing every engineer an AI coding agent is the easy part. Instacart runs one of the most complex B2C and B2B platforms in commerce, connecting customers, shoppers, retailers and brands across more than 2,200 retail banners and about 100,000 stores. Once our engineers had a racecar, we learned the real work was building the road. This talk goes deep on Blueberry, our AI SRE, which investigates every production alert before a human picks it up. It also covers how we're re-architecting our platform so agents can build faster and more safely, with humans in control at every step. Key takeaways: → Start where vigilance beats judgment. Our first attempt, pointing AI at entire incidents, was slow and error-prone. Starting from a single alert, with specific context, is what made our AI SRE useful. → Institutional memory beats a bigger model. Accuracy came from our own incident history and the runbooks and skills each team teaches the agent, not from a newer model. → Match human oversight to the risk. Agents investigate on their own, propose production changes for human approval, and leave novel, high-stakes decisions to people. Autonomy grows only as it's earned. → Humans draw the boundaries. Clear service boundaries, contracts and review loops let agents take on rewrites we once called impossible. → Not less AI, better AI. Route each task to the right model, and give every developer visibility into their own usage