AI eliminates technical debt: Instacart stops worrying about code

What happened?

At the VB Transform 2026 conference, Instacart's CTO, Anirban Kundu, presented a provocative thesis: AI can absorb most of the repetitive work of developers, eliminating technical debt in the process. According to Kundu, in 97% of cases, Instacart developers no longer read AI-generated code. New code is generated or regenerated weekly, and idle systems are automatically discarded. "The benefit is that we no longer worry about technical debt," Kundu said.

Instacart has implemented a system of AI agents that generate code, perform automatic evaluations (about 7,000 per month with 99.9% accuracy), and manage site reliability (SRE) through a system called Blueberry, which monitors Slack channels and detects production incidents. The company has invested years in its Atoms project to break down monoliths into RPC-based architectures, facilitating constant code regeneration. This transformation is not improvised: it responds to a deliberate strategy of migrating from monoliths to microservices, allowing code to be generated and discarded without accumulating debt.

Why is it important?

Technical debt is one of the biggest burdens in software development. Traditionally, companies accumulate legacy code, patches, and workarounds that slow down innovation. Instacart's approach suggests that if code is constantly generated and regenerated by AI, technical debt simply disappears because obsolete code never accumulates. This could transform software project management, reducing maintenance costs and accelerating time-to-market. According to industry data, companies spend between 30% and 50% of their IT budget on maintaining legacy systems; if this strategy became widespread, those resources could be redirected to innovation.

Additionally, the AI-based SRE system (Blueberry) has increased incident detection accuracy from 60% to 90%+, by training on Instacart's own historical data rather than generic data. This demonstrates that AI can surpass human intuition in complex diagnostics. In contrast, traditional monitoring systems like PagerDuty or Datadog often generate false positives that require manual intervention. Blueberry, by learning from Instacart's past incidents, manages to identify patterns that previously went unnoticed.

Consequences for the industry

If Instacart's strategy becomes widespread, the implications are profound:

• Redefinition of the developer role: It will shift from writing code to orchestrating AI agents, defining intentions and handling exceptions. As Kundu said: "At the most basic tactical level, it's no longer about creating code, but navigating around the AI system to get what you want." This resembles the shift from programming in assembly to high-level languages: the programmer focuses on logic, not low-level details.
• Obsolescence of traditional code reviews: When AI generates syntactically correct code, review focuses on intent rather than syntax. Instacart is already training its developers in an "intent model," where they learn to ask the right questions to AI models. This could eliminate the need for pull requests and manual code reviews, saving hours of work.
• Reduction of maintenance costs: Companies could save millions by eliminating accumulated technical debt. For example, a Stripe study estimated that developers lose 17 hours weekly on legacy code; with generative AI, that time is drastically reduced.
• New challenges: Dependence on AI requires robust evaluation systems and strategies to handle the 3% of legacy, compliance, or latency-sensitive code. Instacart still keeps that 3% under human supervision and is working on "smoothing" those parts through the Atoms project. Automatic evaluation is key: the 7,000 monthly evaluations with 99.9% accuracy are a standard that other companies will need to emulate.

What should readers know?

Not everything is immediately applicable. Instacart has invested years in its Atoms project to break down monoliths into RPC-based architectures. Additionally, the 3% of code that still requires human attention includes legacy, compliance, and critical latency systems. The strategy works best in environments where code can be regenerated frequently; it is not a silver bullet for all cases. For example, companies with embedded or real-time systems could hardly rewrite code weekly.

Companies that want to follow this path should prioritize creating automatic evaluation pipelines and train their teams in an "intent model," where developers learn to ask the right questions to AI models. They should also consider the maturity of their architecture: if they still have large monoliths, decomposing into smaller services is a prerequisite. Kundu warned that his team started with monoliths and is migrating to RPC, a process that took years.

"At the most basic tactical level, it's no longer about creating code, but navigating around the AI system to get what you want," Kundu said.

Instacart's case is an early benchmark of how AI can transform software engineering, but each organization must evaluate its own technical and cultural maturity before adopting this approach. History shows that similar changes, like the adoption of cloud computing, required years of adaptation. What seems like an exception today could become the norm in the next decade.