Instacart is challenging the conventional approach by asking if much of the engineering work today should actually be handled by machines. At VB Transform 2026, CTO Anirban Kundu explained that development teams often spend time on repetitive, high-volume tasks that AI agents could manage, freeing humans to focus on areas requiring judgment and intent. Remarkably, Instacart’s engineers now rarely read code directly—97% of the time they work through AI systems.
AI handles most code generation and boilerplate tasks, especially for projects updated weekly. This shift means tech debt is no longer a primary concern; inactive code can simply be dropped and rebuilt as needed. However, about 3% of the work, including legacy and compliance tasks, still demands human oversight. Instacart is addressing these by modularizing their systems through their project Atoms, moving from monolithic to RPC-driven architectures.
Evaluation processes are evolving too. Instead of traditional code reviews, the focus is now on an intent-based model where developers query AI with specific goals. Instacart runs thousands of automatic evaluations monthly with near-perfect accuracy.
Additionally, their AI-powered site reliability engineering system, trained on Instacart’s historic incidents, significantly improves detection and resolution of production issues. For example, their internal AI agent Blueberry identified a subtle production hiccup faster than the human team could, showcasing how AI can overcome typical human intuition limits.
Looking forward, engineers’ roles will center on guiding AI systems, designing evaluations, managing experiments, and handling edge cases. Domain expertise is also being democratized across teams through shared code definitions, moving away from siloed ownership of code changes.
