Editorial illustration for Instacart built an AI system trained on years of its own incident data
Instacart's AI Agents Write 97% of Code Unseen
Instacart built an AI system trained on years of its own incident data
Instacart's engineers have stopped reading most of their own code. CTO Anirban Kundu told the audience at VB Transform 2026 that in 97% of cases, the company's builders no longer look at what AI agents generate. That's not a productivity stat pulled from a slide deck. It's a description of how Instacart now writes software.
Kundu's argument runs deeper than the usual AI-coding-assistant pitch. He says the repetitive, high-volume work that eats up engineering hours, boilerplate, regeneration, routine maintenance, should belong to machines, freeing humans for the calls that actually need judgment. Instacart has built its own AI system for this, trained on years of internal incident data, and it's already changing how the company thinks about problems like tech debt. Code that isn't active gets dropped and rebuilt rather than patched, an approach Kundu compares to how engineers once treated assembly or object code: disposable, regenerable, not precious.
The remaining 3%, legacy systems, compliance-heavy workflows, anything latency-sensitive, still needs a human eye. But Kundu's framing raises a bigger question for engineering teams everywhere: how much of what developers do today is actually necessary, and how much is just habit.
“The benefit of that is we don't care about tech debt anymore,” Kundu said. “Things that are not active just get dropped out and then it gets rebuilt, kind of like how we used to build assembly code or object code.”
Why this matters
Kundu's 97% figure is the kind of number that should make engineering leaders pause rather than nod along. Training an SRE agent on your own incident history is a smarter move than bolting on a generic anomaly detector, and it fits a pattern we've seen at companies willing to treat operational data as a real asset rather than log-file exhaust. But "builders don't read code anymore" is a claim that deserves scrutiny, not applause.
Who verifies the agent's root-cause calls when an incident actually escalates? What happens when the training data reflects Instacart's specific infrastructure quirks and the system meets a failure mode it's never seen? For founders and platform teams watching this, the lesson isn't "replace your SREs." It's that domain-specific incident data, accumulated over years, is worth more than another vendor's off-the-shelf model.
The harder question Instacart hasn't answered yet: what's the failure rate when human intuition would have caught something the model missed?
Common Questions Answered
What percentage of code generated by Instacart's AI agents do engineers no longer review?
According to Instacart's CTO Anirban Kundu, engineers stop reading code in 97% of cases where AI agents generate it. This represents a fundamental shift in how Instacart's builders approach software development, moving away from traditional code review processes for the vast majority of AI-generated work.
How did Instacart train its AI system to improve incident response?
Instacart trained an SRE agent specifically on the company's own incident history data accumulated over years of operations. This custom approach is more effective than using generic anomaly detectors because it learns from Instacart's unique operational patterns and historical issues.
What does Instacart's CTO claim about tech debt since implementing AI code generation?
CTO Anirban Kundu stated that Instacart no longer needs to worry about tech debt because inactive code simply gets dropped and rebuilt by AI agents as needed. This approach mirrors how assembly code and object code were traditionally managed, allowing the system to continuously refresh itself without accumulating legacy technical obligations.
What types of engineering work does Instacart's AI system handle according to the article?
Instacart's AI system handles repetitive, high-volume work that typically consumes significant engineering hours, including boilerplate code and code regeneration tasks. By automating these routine aspects of software development, the company has fundamentally changed how its builders allocate their time and effort.
Further Reading
- Instacart's CTO says AI made the company stop worrying about tech debt - VentureBeat
- Instacart's AI Pricing May Be Inflating Your Grocery Bill - Consumer Reports
- Instacart Ends AI-Driven Price Experiments After FTC Criticism - Fintool
- OpenAI's Operator Agent Reportedly Executed Unauthorized $31.43 Transaction Despite Safety Protocol - AI Incident Database
- Instacart ends AI-driven price experiments after criticism - Reuters