Stay connected with KayaToday, follow us on Instagram and Facebook for the latest news and reviews delivered straight to you.
For decades, tech debt has been one of the most persistent headaches in software engineering: the accumulated mess of legacy code, half-finished systems, and outdated architecture that slows teams down and quietly drains engineering capacity. Instacart’s CTO thinks AI has just made that problem largely irrelevant, and the reasoning is more substantive than it first sounds.
Speaking at VB Transform 2026, Anirban Kundu laid out how Instacart has restructured its engineering organisation around AI agents, to the point where its developers no longer read code in 97% of cases. The claim is striking, but the mechanics behind it reveal something genuinely significant about where software development is heading.
When Code Becomes Disposable, Debt Stops Accumulating
The core insight Kundu offered is that tech debt only matters when code persists. In Instacart’s current setup, AI agents handle the bulk of code generation and boilerplate, particularly for newer projects where code is generated or regenerated on a weekly basis. Because the code is rebuilt so frequently, the old accumulation of cruft simply does not survive long enough to become a liability.
“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.”
The analogy to assembly code is instructive. Nobody worries about maintaining compiled machine instructions because they are outputs, not assets. Kundu is arguing that AI-generated code is moving toward the same status: something produced on demand rather than something carefully maintained over years.
That said, the 97% figure has a meaningful asterisk. The remaining 3% covers legacy systems, compliance-sensitive workflows, and latency-critical infrastructure, along with large volumes of dead or partially active code that still requires careful human judgment. Instacart is working through this through an internal initiative called Atoms, which breaks monolithic systems into modular components and rebuilds them on cleaner, remote procedure call-driven architectures. The goal is to gradually shrink that 3% rather than pretend it does not exist.
Evaluation Is the New Code Review
Shifting code generation to AI does not eliminate the need for rigorous quality control. It just changes what quality control looks like. Traditional code review becomes less meaningful when the syntax and structure are reliably correct by default. As Kundu noted, “the lines of code are going to be correct, the syntax is going to meet your expectations,” so the question shifts from whether the code is well-written to whether it actually does what was intended.
Instacart’s answer is an intent-based evaluation model. Engineers are trained to ask AI systems the right questions from an intent perspective, and the system is then evaluated independently on whether it delivered the right outcomes. The scale is substantial: roughly 7,000 automatic evaluations run each month, and the system handles more than 8,000 real-time developer queries with approximately 99.9% accuracy.
This is a meaningful reorientation of what engineering skill looks like. The tactical work is no longer writing or reviewing code. As Kundu put it, “the most tactical level going forward is going to be, how do you navigate around the AI system to give you what you want?” Designing evaluation processes, coordinating experiments, managing constraints like limited test traffic, and identifying edge cases before they become incidents: these are the skills that matter now.
The AI Colleague That Caught What the Humans Missed
Perhaps the most concrete illustration Kundu offered was Instacart’s agentic site reliability engineering system, built around an internal tool called Blueberry. Rather than training the system on generic failure patterns, Instacart fed it years of its own incident history and root-cause analyses, giving it a highly specific model of how Instacart’s infrastructure tends to break and how engineers have historically diagnosed and fixed those breaks.
The results are measurable. Accuracy in detecting and mitigating production issues has risen from roughly 60% to more than 90%. Blueberry monitors around 200 Slack channels simultaneously, watching for patterns across human conversations and system alerts.
Kundu described one incident involving a database shard backed by an Amazon EBS volume that experienced a brief disruption. The human team was scrambling to identify the cause and had not yet suspected AWS disk issues. About 20 minutes into the incident, Blueberry posted on Slack, identifying a specific anomaly and connecting it to a feature-flag-like system called “roulette” that had been rolling out changes at too aggressive a cadence.
“Would have a human been as quick? I think the problem is human intuition would hold us back a little bit,” Kundu said. Humans default to patterns they have encountered before, then resort to systematic debugging. Blueberry, by contrast, scans everything simultaneously without anchoring on prior assumptions, which makes it more comprehensive even if it lacks the contextual judgment that experienced engineers bring to genuinely novel situations.
What This Means for Engineering Teams Beyond Silicon Valley
The implications of Instacart’s model extend well beyond one American grocery delivery company. For technology teams across Southeast Asia, including the fast-growing engineering organisations at Malaysian and Singaporean firms navigating their own digital transformation pressures, the shift Kundu describes points to a genuine restructuring of how software talent should be developed and deployed.
Domain knowledge is being decoupled from code ownership. Instacart is moving away from a model where a single team controls a codebase and bottlenecks all changes through that ownership. Instead, domain knowledge is being encoded into definitions and specifications that any team can access and use. “We’re trying to move into a world where the code becomes completely democratised across groups,” Kundu said.
That democratisation has real consequences for how engineering organisations should hire, train, and structure themselves. The premium is shifting away from deep implementation skill and toward the ability to reason about systems, define intent clearly, design evaluations, and handle the exceptions that AI cannot yet manage reliably. For businesses and engineering leaders in this region watching these trends, the question is no longer whether AI will change software development. It is how quickly to restructure teams around the assumption that it already has.
