How Instacart's Blueberry AI Rethinks On-Call Troubleshooting for Web Teams

A Familiar Scene for Web Teams

It's 2 a.m. Your site's product page is throwing errors. The on-call engineer is awake, staring at a dashboard, trying to piece together what changed. They check deployments, skim logs, search docs, and compare symptoms to past incidents. For many web teams, that's the hardest part of incident response—not the fix, but the context gathering.

Instacart knows this pain. Their engineering team built an AI assistant called Blueberry to help on-call engineers handle production incidents faster. It's not a magic wand—it's a system that gathers context, forms hypotheses, and hands the engineer a head start. And the results are hard to ignore: in April, Blueberry ran about 25,000 diagnostics across 270 Slack channels, with accuracy jumping from 60% to over 90%.

What Blueberry Actually Does

Blueberry plugs into Slack, where many incident response workflows already live. When an alert fires, it spins up about ten sub-agents in parallel. These agents pull info from internal sources—incident history, service ownership data, logs, deployments, and other debugging signals. Within about three minutes, they produce an initial diagnosis with evidence, right inside the Slack thread where the engineer is already working.

That's a big deal. Instead of leaving the chat to chase down context, the engineer gets a synthesized picture. The system doesn't make changes to production—it assists the human. The engineer still owns the diagnosis, the decision, and the fix.

Why Web Design Teams Should Care

You might be thinking, "That's for backend SREs, not for me." But think about your own workflows. When a client reports a broken layout or a slow page, how long do you spend figuring out what changed? Maybe it's a CMS update, a new plugin, or a misconfigured CDN. The same principle applies: context is everything.

Blueberry's approach shows that AI can help with the tedious parts of troubleshooting—not by replacing the designer or developer, but by putting the right information in front of them. That's a lesson for any web design team that's ever lost an afternoon to a mystery bug.

Learning from 14 Years of Incident History

One of the most interesting parts of Blueberry is its reliance on history. Instacart fed the system over 14 years of incident data. That's not just a pile of old reports—it's a knowledge base that helps the AI recognize patterns. When a new incident looks like one from 2019, Blueberry can surface that connection.

For web teams, that's a nudge to keep better records. Do you document every bug you fix? Every weird CSS quirk? Every time a plugin update broke the site? Probably not. But that kind of history is gold. It's the raw material that makes any AI assistant—or even a well-organized wiki—useful.

Tools That Talk to Each Other

Blueberry doesn't rely on a general-purpose language model. It connects to specific tools and data sources. The architecture uses a persistent, tool-aware framework, with an MCP (Model Context Protocol) catalog and a sidecar setup. That means the AI can call the same tools an engineer would: query logs, check deployment status, pull service ownership info.

This is a crucial point for web design teams. An AI that only knows generic web design principles is limited. But an AI that can check your site's uptime, look at recent commits, or scan your design system's changelog? That's powerful. The key is integration—making your AI assistant aware of your actual stack.

Not a Replacement, a Multiplier

Instacart's VP of Engineering, Siby Alappatt, calls Blueberry a "force multiplier." It doesn't replace the on-call engineer; it makes them better. By handling the initial information gathering, Blueberry lets engineers jump straight to analysis and action.

That's a mindset shift worth adopting. Instead of fearing AI as a job killer, think of it as a teammate that handles the boring stuff. For web designers, that could mean an AI that scans your site for accessibility issues before a client points them out, or one that flags a performance regression in a pull request. The goal isn't to remove the human—it's to give them superpowers.

What's Under the Hood

Blueberry's architecture is worth a closer look. It uses a persistent state to track progress across steps. That means if the AI is investigating a hypothesis and needs to check another log, it remembers what it was doing. It also has a review and improvement loop, so the system learns from each incident.

In a month, Blueberry handled roughly 25,000 diagnostics, with a workflow success rate of 99.9%. It made over 58,000 MCP tool calls and adapted to about 60 different team workflows. Those numbers suggest the approach isn't just a demo—it's a real, working system.

Practical Takeaways for Web Design Teams

You don't need to build a full AI ops platform to learn from Blueberry. Here are a few things you can start doing today:

• Document everything. Keep a searchable log of bugs, fixes, and odd behavior. Future you will thank you.
• Integrate your tools. Make sure your design handoff, CMS, and monitoring tools can talk to each other. Automate the boring parts.
• Use AI for context, not decisions. Let AI gather data and suggest possibilities, but keep a human in the loop for final calls.
• Build a feedback loop. After every incident, ask what worked and what didn't. Feed that back into your process.

Blueberry is a reminder that the biggest wins in web design often come from improving the environment around the work, not just the pixels on the screen. By giving your team better context and faster answers, you can turn a stressful 2 a.m. call into a manageable problem.

The Future of AI in Web Operations

Instacart's Blueberry is one of the first large-scale examples of AI driving incident response. But it won't be the last. As AI tools get better at understanding context—and as more platforms expose APIs for logs, metrics, and deployment data—we'll see more teams adopting similar approaches.

For web design teams, the opportunity is clear. Whether you're a solo freelancer or part of a large team, there's room to use AI to cut down on the time you spend hunting for information. The goal isn't to automate away the craft—it's to make room for more of it.