How Blueberry, Instacart's AI, Speeds Up On-Call Firefighting

The Kitchen Fire Analogy

You're running a busy restaurant. Orders pile up, the grill's smoking, and then the ticket machine jams. No time to dig through the recipe binder or call the supplier—you need to fix it now. That's the daily reality for on-call engineers at a big tech company like Instacart. Every minute of downtime costs money and trust.

The old way of handling these crises was painfully slow. An engineer had to pull context from a dozen sources: Which service just failed? Who owns it? What changed in the last deployment? Are there logs or metrics? Is this a known issue from six months ago? That digging could take twenty minutes or more before actual troubleshooting even began.

Instacart built an AI assistant called Blueberry to change that. It's not a magical fix-all. It's more like a sous-chef who preps all the ingredients while the head chef focuses on the cooking. Blueberry collects the background info, forms hypotheses, and hands the engineer a head start.

What Blueberry Actually Does

Blueberry is an AI-powered incident response system that lives inside Slack, where Instacart's engineering teams already talk during outages. When an alert fires, Blueberry spins up about ten sub-agents in parallel. These agents dig into internal data—incident history, service ownership, logs, deployment records—and post their findings directly into the Slack thread. The whole process takes about three minutes to produce an initial diagnosis.

Three minutes. That's a huge time saver. Instead of starting from zero, the engineer gets a summary of what's happening, what changed, and likely root cause. Blueberry never touches the production environment; it only provides information. The human engineer still makes the final call on what to fix and how.

Better Than Guesswork: The Data Behind It

What makes Blueberry more than a fancy chatbot? It's access to history. Instacart has over 14 years of incident records baked into the system. When Blueberry suggests a root cause, it's not guessing from a generic model—it's comparing current symptoms against thousands of past incidents to find patterns.

That approach paid off. In April alone, Blueberry handled about 25,000 diagnostic sessions across more than 270 Slack channels. Diagnostic accuracy jumped from 60% to over 90% once historical data was fully integrated. That's the difference between a cook who guesses at seasoning and one who remembers exactly what worked last time.

How It Fits Into the Kitchen Workflow

One of the best parts of Blueberry is that it lives where the engineers already are. No separate dashboard to check, no tool to switch to. The AI posts its findings right in the Slack thread, so the team keeps their conversation flowing without context switching. It's like having a prep cook who slides you the chopped veggies without asking you to leave the stove.

This integration extends beyond the immediate incident. Blueberry also captures knowledge gained during each troubleshooting session. It stores decisions made, fixes applied, and context around them. Next time a similar issue crops up, the system can reference what worked before—even for teams that weren't part of the original incident.

The Recipe for Success: Tools and Feedback

Building an AI that actually helps in production is tricky. The core challenge is keeping the AI grounded in facts, not just plausible-sounding guesses. Blueberry uses a tool-calling pattern where agents fetch data from connected systems as they work. They also keep a persistent state of findings, so they don't lose track of what they've already investigated.

Engineers can correct or validate Blueberry's suggestions, and the system learns from those interactions. It's not a one-and-done AI; it improves over time.

After a month of operation, Blueberry had a workflow success rate of 99.9%, made over 58,000 MCP tool calls, and adapted to the working styles of about 60 different teams. Those numbers suggest it's not just a toy but a reliable part of incident response.

What This Means for Cooking and Beyond

You might be wondering what an AI incident response tool has to do with cooking. Well, think about a professional kitchen. It's high-pressure, and speed and accuracy matter. A good sous-chef doesn't just follow orders; they anticipate needs, know where everything is, and can spot a problem before it becomes a disaster. Blueberry is like that sous-chef for engineers.

For cooking enthusiasts, the lesson is about preparation and knowledge. When you're in the middle of a complex recipe, you don't want to stop and Google every step. You want a well-organized mise en place, a clear understanding of your ingredients, and maybe a few notes from the last time you made it. Blueberry brings that same philosophy to software operations.

The Future of AI in the Kitchen

Instacart's Blueberry shows that AI can be a powerful ally in fast-moving environments. It doesn't replace the human; it amplifies their abilities. The same could apply to cooking. Imagine an AI that knows your pantry, your past recipes, and your dinner guests' preferences. It could suggest substitutions, time-saving techniques, or even warn you that your oven runs hot.

That's not science fiction. It's the same pattern as Blueberry: combine a smart model with your personal history and current context, then present the information in a way that helps you make better decisions faster.

Lessons for Home Cooks and Engineers Alike

So what can we take away from Instacart's experiment? First, context is king. The more relevant information you have at your fingertips, the quicker you can solve problems. Second, don't be afraid to let AI do the prep work. You still make the final decisions, but you can save time by letting an assistant gather the facts.

Finally, remember that tools are only as good as their data. Blueberry's accuracy improved dramatically when it had 14 years of incidents to learn from. In cooking, that translates to keeping a journal of your own experiments—what worked, what flopped, and why. Over time, that's your personal incident history.

Whether you're debugging a server or perfecting a sauce, the principles are the same: gather context, use past wisdom, and act decisively. Blueberry is just one example of how AI can help us do that better.