Instacart CTO: AI now does 97% of code work, engineers focus on judgment

At VB Transform 2026, Instacart CTO Anirban Kundu outlined a provocative reimagining of software engineering: the majority of the work engineers do today—particularly the draining, repetitive, high-volume tasks—should be performed by AI agents instead. This shift would liberate human developers to focus on problems that require judgment, intent, and exception handling.

The scale of this change at Instacart is striking. In 97% of cases, the company's builders do not read code anymore. AI agents handle the bulk of code generation and boilerplate work, especially on newer projects where code is generated or regenerated on a weekly basis. Kundu emphasized that this does not mean code is never reviewed by humans—agents perform "pretty serious evals"—but the volume of human involvement has shrunk dramatically.

The most tangible benefit, Kundu argued, is the elimination of tech debt as an engineering concern. Tech debt—the accumulated burden of maintaining older, poorly-organized code that slows development—has historically consumed significant engineering resources. By allowing AI to constantly regenerate code rather than patch and maintain legacy systems, Instacart sidesteps this problem entirely. "We don't care about tech debt anymore," Kundu said.

This represents a fundamental reframing of the engineering role. Kundu noted: "In the past, the tactical level was the creation of the code. 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?'" Rather than writing and maintaining code, engineers increasingly manage the prompts, requirements, and workflows that guide AI-generated code, and focus on the judgment-heavy decisions that machines cannot yet handle.

Instacart's approach reflects a broader shift in how AI is reshaping software engineering workflows. Rather than treating AI as a tool that augments human coders, the company has reorganized its dev process around the assumption that machines should handle the repetitive, high-volume tactical work—code generation, boilerplate, and routine maintenance. This frees humans to engage in activities that machines cannot easily replicate: architectural decisions, exception handling, and judgment calls.

The 97% figure Kundu cited is striking not because it means code is never reviewed, but because it signals a fundamental inversion of engineering priorities. Historically, developers spent significant time reading, debugging, and refactoring existing code. Instacart's model suggests that cost is now externalized to AI, which regenerates code on a weekly basis in newer projects. This has a knock-on effect: tech debt, which accumulates when codebases become too complex or poorly organized to maintain efficiently, simply ceases to be a drag on the team's velocity. The company no longer needs to allocate engineering hours to debt reduction because the system is constantly regenerating rather than patching.

No comments yet. Be the first to share your thoughts!