The Death of Tech Debt
For decades, technical debt has been the quiet, gnawing anxiety at the heart of every engineering team. It’s that pile of "temporary" fixes, hacked-together workarounds, and messy modules we promise ourselves we’ll address "later." But at VB Transform 2026, Instacart CTO Anirban Kundu introduced a provocative new perspective: What if you just stopped worrying about it entirely?
Kundu’s argument isn’t that debt doesn't exist, but rather that our traditional obsession with cleaning it up is fundamentally mismatched with an AI-driven future. If an AI agent can generate, test, and deploy functional code in milliseconds, the long-term maintainability of that code starts to matter less. Instead of painstakingly refactoring, you might just throw the old code away and build it again.
It’s a audacious shift, and for Instacart, it’s already becoming reality.
The 97% Solution: Code as Ephemeral Material
The numbers Kundu shared were staggering: in 97% of cases at Instacart, the builders don't even read the code they’re deploying. AI agents generate the bulk of the logic and boilerplate. For many projects, the code is regenerated from scratch on a weekly basis.
"The benefit of that is we don't care about tech debt anymore," Kundu said during his talk. "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."
This effectively turns code from a durable asset into ephemeral material. In this model, the "truth" isn't a long-lived, carefully curated repo; the truth is the intent—the prompt or requirement—you feed to the system. The agents ensure the syntax is correct and the functionality meets the spec. For any code that’s not mission-critical, the idea of carrying technical debt for six months just to save a few hours of development time suddenly feels like archaic thinking.
Rethinking Reliability: Blueberry and the Agentic SRE
While the generative aspect of Instacart’s strategy is eye-catching, the shift in how they handle site reliability engineering (SRE) might be more consequential. Instacart has built an agentic SRE system nicknamed "Blueberry," helping drive the broader industry trend toward AI-powered SRE workflows. It's not learning from generic, textbook examples of failure.
Instead, they’ve trained the system on an extensive backlog of their own incidents and post-mortem analyses. It is a system built in the crucible of their specific architectural failures.
The results have been dramatic. Since implementing Bluebery, the accuracy in detecting and mitigating production issues has jumped from roughly 60% to over 90%. In one instance, while human engineers were struggling to identify the cause of a database shard issue, Blueberry analyzed the patterns, made sense of a specific "hiccup," and traced it back to an underlying feature-flag system that had been misconfigured. Bluebery saved the team from spiraling down a typical human path of trial-and-error debugging.
Kundu’s insight here is crucial: human intuition is valuable, but it's often constrained by patterns it's seen before. Autonomous agents can look at the entirety of a system’s behavior in real-time, untroubled by our emotional biases or limited working memories.
The New Engineering Blueprint: Intent, Not Syntax
If agents are writing the code and agents are managing the reliability, what on earth are the engineers doing?
Kundu argues the role of the modern developer is fundamentally changing. The most tactical work for an engineer is no longer pounding out lines of code; it’s about navigating the AI system. Engineers must now focus on supervising evaluations, managing constraints, and—perhaps most importantly—defining the intent behind the features being built.
Beyond that, Instacart is tackling the issue of domain knowledge bottlenecks. In traditional engineering structures, specific teams "own" specific parts of the codebase, creating slow, siloed workflows. They are trying to move toward a model where domain knowledge is embedded into specifications and definitions that are democratized across all engineering groups.
A Harder Look at the Remaining 3%
Of course, not everything can be thrown away and rebuilt on a whim. Kundu acknowledged that 3% of Instacart’s environment—encompassing legacy systems, compliance requirements, and latency-sensitive workflows—still requires precise, human-led attention.
These areas are the "Atoms" project’s focus: the team is methodically breaking down monoliths and rebuilding them into a cleaner, modular RPC-driven architecture. This is a deliberate, high-stakes endeavor that sits in sharp contrast to the rapid, disposable nature of the other 97% of their work.
Is This the New Normal?
Instacart's approach is fascinating because it treats AI not just as a productivity tool, but as a mechanism to fundamentally alter the cost-benefit analysis of infrastructure. When regeneration is cheaper than maintenance, technical debt stops being a debt and starts looking like a relic.
Still, this path isn't for everyone. It requires a sophisticated evaluation system—the kind that can run 7,000 automatic tests a month and provide 99.9% accuracy—to keep the system from drowning in AI-generated chaos.
For now, Instacart is showing us a future where developers spend less time fixing old code and more time defining what the next version should actually be. Whether the changing landscape of SRE allows the rest of the industry to follow them into that future, or if we'll find ourselves overwhelmed by the complexity of our own autonomous systems, remains to be seen. But one thing is clear: the conventional wisdom about technical debt just got a lot less certain.
Source: VentureBeat - Instacart's CTO says AI made the company stop worrying about tech debt
