Instacart's CTO Says Engineers No Longer Read 97% of the Company's Code

Instacart's chief technology officer says his engineers have mostly stopped reading code. It's the strategy.

Speaking at VB Transform 2026, CTO Anirban Kundu said AI agents now handle roughly 97% of the coding work at Instacart, according to VentureBeat. Human engineers, in his telling, have shifted from writing code to managing the AI systems that write it for them.

"In the past, the tactical level was the creation of the code," Kundu said, per VentureBeat. "In the most tactical level going forward, it's going to be, 'How do you navigate around the AI system to give you what you want?'"

The Tech Debt Argument

Kundu's boldest claim: Instacart doesn't worry about technical debt anymore. When code isn't being used, it gets dropped and rebuilt from scratch by AI, rather than patched and maintained the old way.

"The benefit of that is we don't care about tech debt anymore," Kundu said, according to VentureBeat. He compared it to how engineers stopped hand-writing assembly code once compilers got good enough to do it for them.

This represents a verifiable shift in how software gets built if it holds up. Tech debt, the accumulated cost of quick fixes and outdated code that eventually slows a company down, has been an industry-wide headache for decades. If AI can genuinely regenerate systems on demand instead of accumulating cruft, that's a meaningful efficiency gain, not just a talking point for a stage.

Where Humans Still Matter

Instacart isn't claiming full automation. Kundu said the remaining 3% of engineering work still requires human hands: legacy systems, compliance-driven code, and latency-sensitive workflows where mistakes are expensive.

The company has a project called Atoms specifically for this. Engineers are breaking down older "monolith" systems and rebuilding them on remote procedure call (RPC) architecture, according to VentureBeat. That's slow, deliberate human work, the opposite of the AI-driven speed Kundu describes elsewhere.

That distinction matters. Kundu isn't saying AI replaced engineers. He's saying AI absorbed the repetitive, high-volume work, freeing humans for judgment calls, exceptions, and the risky legacy systems nobody wants to touch with a fully automated tool.

The Evaluation Problem

If nobody's reading the code, how do you know it's right?

Kundu acknowledged this directly. Traditional code review doesn't work the same way when AI is generating the code, because as he put it, "the lines of code are going to be correct, the syntax is going to meet your expectations," according to VentureBeat. Syntax being correct isn't the same as the code doing what the business actually needs.

Instacart's answer is what Kundu called an "intent model," training engineers to ask AI systems the right questions rather than checking the output line by line. The company says it runs roughly 7,000 automated evaluations every month and its system fields more than 8,000 real-time developer queries at about 99.9% accuracy, per VentureBeat.

That's an internal, self-reported number. There's no independent audit cited in the available reporting confirming that 99.9% figure, and Instacart hasn't published the methodology behind it. It's a legitimate figure to report, but readers should know it comes from the company measuring itself.

The SRE Angle

Instacart also built an AI site-reliability engineering (SRE) tool, trained specifically on the company's own history of outages and fixes rather than generic industry failure data, according to VentureBeat. The tool, called Blueberry, monitors more than 200 Slack channels and looks for patterns in human conversations and system alerts to catch problems early.

Kundu said accuracy in detecting and mitigating production issues jumped from around 60% to more than 90% after deploying the tool. Again, that's a company-reported figure without independent verification available in current reporting.

What This Actually Signals

The skeptical read here isn't that Kundu is lying. It's that a CTO on stage at an industry conference has every incentive to describe his company's AI adoption in the most impressive terms possible. VB Transform is a marketing moment as much as a technical disclosure.

The fair counterpoint: engineers not reading code isn't inherently reckless if the evaluation layer genuinely catches errors before they hit production, and Instacart's approach of training its SRE model on its own incident history, rather than generic data, is a more rigorous method than most companies attempt.

What's unresolved is how this holds up outside a keynote. Instacart hasn't published a technical paper or third-party audit on the 99.9% accuracy or the 60-to-90% SRE improvement. Whether other engineering leaders adopt this model, or whether Instacart's own systems face a major outage that tests the AI SRE tool under real pressure, will tell you more than any conference talk can.