Why Instacart's Blueberry Bot Feels Like a Creative Partner in an Outage

When you hear "creative arts and crafts," you might think of a painter at an easel, a potter at a wheel, or a kid with a glue stick and a pile of macaroni. But creativity isn't just for the studio. It's also what happens when a team of engineers stares down a production outage at 3 a.m. and has to improvise a fix before the coffee kicks in. That's the world of incident response, and it's about to get a whole lot more interesting.

Instacart, the grocery delivery giant, built an AI assistant named Blueberry to help its on-call engineers handle production faults. Blueberry isn't a chatbot that spits out generic advice. It's a system that taps into years of incident history, live operational data, and a fleet of sub-agents to give engineers a head start when things go sideways. The result? A tool that feels less like a search engine and more like a seasoned partner who's seen it all before.

The Pain of Starting from Scratch

Anyone who's ever been on call knows the drill. An alert fires, and suddenly you're scrambling to piece together what's happening. You check which service owns the failing component, review recent deploys, dig through logs, scan metrics, search for relevant docs, and try to remember if you've seen this symptom before. It's a lot of context-gathering before you can even start diagnosing the actual problem.

That's exactly the pain Blueberry was built to address. By automating the busywork, it lets engineers jump straight into the detective work. Instead of spending twenty minutes just figuring out where to look, they can start with a list of likely culprits and a stack of relevant data already pulled together.

Blueberry's Recipe: A Dash of AI, a Pinch of History

So how does Blueberry actually work? It's not just a single AI model. It's a system that spins up around ten sub-agents in parallel whenever an alert triggers. These agents fan out across Instacart's internal resources—incident histories, service ownership maps, logs, deployment records, and other debugging signals—to gather context and generate hypotheses.

The key ingredient is the fact base. Instacart has over fourteen years of incident data, and Blueberry uses that history as a grounding source. The system doesn't just guess; it compares current symptoms to past incidents and surfaces patterns that a human might miss. That's why its diagnostic accuracy jumped from 60% to over 90% in a matter of weeks.

Working Where the Work Happens

One of the smartest moves was integrating Blueberry directly into Slack. Engineers at Instacart handle incidents through Slack channels, so Blueberry lives right there in the thread. You don't have to switch to a separate dashboard or learn a new interface. The bot posts its findings, updates its theories, and even asks clarifying questions—all without leaving the conversation.

It's a bit like having a tireless intern who's read every postmortem ever written and can fetch any log you need, without ever getting tired or making a snarky comment. The system collects information, generates hypotheses, and assists with debugging, but it never makes changes to the production environment on its own. That's left to the humans, which is probably a good thing.

The Numbers Behind the Magic

Blueberry isn't just a fun experiment. In April alone, it completed about 25,000 diagnoses across over 270 Slack channels. That's a lot of incident response. The workflow success rate hit 99.9%, and the system made more than 58,000 tool calls to internal resources. It even adapts to the working styles of about 60 different teams.

Those numbers are impressive, but what do they mean in practice? For one, they suggest that AI can be a genuine force multiplier for on-call teams. Instead of burning out engineers with repetitive context-gathering, Blueberry handles the grunt work and lets them focus on the creative problem-solving that actually moves the needle.

Learning from Every Incident

One of the coolest parts of the Blueberry system is that it learns from every incident. It doesn't just diagnose and move on. It captures the knowledge that emerges during the resolution process, integrating past cases with team-specific business context. That means the next time a similar issue pops up, the system has even more information to work with.

It's like building a living library of operational wisdom. Every mishap, every late-night fix, every clever workaround becomes part of the collective memory. Over time, that makes the entire organization better at responding to problems, and it reduces the dreaded 'tribal knowledge' problem where only one person knows how to fix a particular issue.

From Zero to Context in Three Minutes

Time is everything in an outage. Blueberry typically produces its initial findings in about three minutes. That's not just fast; it's transformative. Instead of starting from zero, engineers start with a stack of context: relevant logs, deployment info, related systems, and a list of probable causes. It changes the starting line, as Instacart's software engineering director, Alan Wong, put it.

Engineers no longer have to spend precious minutes just figuring out where to look. They can dive straight into verification and remediation. It's a bit like being handed a map before you enter a maze instead of having to feel your way along the walls.

The Framework Behind the AI

Instacart's experience with Blueberry highlights something important: effective AI in operations depends as much on the surrounding framework as on the model itself. The system uses a tool-aware runtime, meaning the agents can call out to connected systems to fetch data as needed. It also maintains persistent state, so it remembers what it's already checked and what it's still investigating.

There's also a review-and-improve loop. The system gets feedback from engineers and uses that to refine its future recommendations. This continuous learning cycle is what takes it from a parlor trick to a genuinely useful tool.

Creativity in the Machine

So where does the 'creative arts and crafts' angle come in? Think about what a good craftsperson does. They take raw materials—wood, paint, fabric—and combine them in novel ways to create something useful and beautiful. Blueberry does the same with data. It takes logs, metrics, and historical incidents, mixes them with a bit of AI, and produces something new: a diagnosis, a hypothesis, a path forward.

It's not creating a painting or a sculpture, but it is creating knowledge. And that's a creative act. The engineers who use it are also engaged in a kind of creative work, piecing together clues and crafting solutions. Blueberry just gives them better materials to work with.

Lessons for the Rest of Us

Whether you're a crafter, a coder, or a bit of both, there's something to learn from Blueberry's success. First, context matters. You can't create anything meaningful without understanding the materials at hand. Second, history is a treasure trove. Past projects, past mistakes, past successes—they're all raw material for future work. Third, tools should fit seamlessly into your workflow, not force you to adapt to them.

And finally, never underestimate the power of a good assistant. Whether it's an AI bot or a helpful colleague, having someone (or something) to handle the tedious parts frees you up to do the creative thinking. That's what Blueberry does for Instacart, and it's a model that could inspire any field.

So the next time you're facing a blank canvas or a blinking cursor, remember: the creative process isn't just about the moment of inspiration. It's about the preparation, the tools, and the knowledge you bring to the table. Blueberry is a reminder that even in the most technical corners of the world, creativity and craftsmanship still rule the day.