Instacart Engineers Read Code Only 3% of the Time Now

The New Reality of Software Work

In 97% of development scenarios at Instacart, engineers no longer read code. The number is striking not because it represents a complete automation of software work, but because it signals a fundamental shift in what engineering teams actually do all day. The grocery delivery platform has restructured its entire development practice around a principle that would have seemed radical five years ago: most of what developers traditionally did should be machine work.

At a recent industry event, Anirban Kundu, Instacart's CTO, laid out a thesis that challenges conventional wisdom about technical skill. The tactical work of software development, he argued, has moved from writing lines of code to navigating AI systems that generate and maintain those lines. For newer projects at Instacart, code is now regenerated weekly, and the company has stopped worrying about technical debt in the traditional sense. If code isn't actively used, it gets dropped and rebuilt when needed, much like compiled binaries in older computing paradigms.

The shift isn't total. Legacy systems, compliance-critical workflows, and latency-sensitive infrastructure still demand human attention. A significant portion of Instacart's codebase consists of what Kundu described as dead, half-active, or deeply entangled logic that AI can't yet safely refactor. But even here, the company is methodically breaking down monolithic systems through an initiative called Atoms, transitioning to remote procedure call architectures that are more amenable to machine-driven generation.

Evaluation Over Syntax

The immediate consequence of machine-generated code is that traditional code review becomes less relevant. Syntax will be correct, style guides will be followed, and boilerplate will be consistent. The real question is whether the code does what the engineer intended, and this requires a different kind of oversight.

Instacart runs roughly 7,000 automated evaluations each month to verify that AI-generated code meets functional expectations. The system also handles more than 8,000 real-time developer queries with accuracy exceeding 99.9%, according to Kundu. The focus has shifted from line-by-line inspection to training developers to ask the right questions of the models, framing requests in terms of intent rather than implementation detail.

This is not a minor adjustment. It requires engineers to articulate what they want the system to do, then validate that the output matches that specification, without necessarily understanding every intermediate step. The skill set tilts toward systems thinking, constraint management, and edge-case identification rather than syntax mastery or algorithmic fluency.

A Site Reliability Agent That Learns From History

Instacart has also applied agentic AI to site reliability engineering, an area where human intuition and pattern recognition have historically been critical. The company trained an SRE agent on years of internal incident reports and root-cause analyses, rather than generic failure scenarios. The result is a system that understands how Instacart's infrastructure specifically tends to break, and how past teams have diagnosed and resolved those issues.

The accuracy of detecting and mitigating production problems has climbed from around 60% to more than 90%, Kundu said. One internal tool, Blueberry, monitors over 200 Slack channels, scanning conversations and alerts for anomalies. During one incident, a database shard experienced intermittent issues tied to an EBS volume disruption. The human team scrambled to diagnose the problem, initially missing the connection to underlying storage behavior.

Twenty minutes into the incident, Blueberry posted a message in Slack identifying the specific infrastructure blip and linking it to a feature-rollout system called roulette that had been pushing changes too aggressively. The team resolved the issue based on that insight. Kundu noted that human intuition might have delayed the diagnosis, because engineers tend to default to patterns they've seen before and then methodically debug. The agentic SRE, by contrast, evaluates a broader set of signals simultaneously and identifies correlations that don't fit familiar narratives.

Democratizing Domain Knowledge

The structural implication of this shift is that domain expertise can no longer be bottlenecked in a single team that owns a particular codebase. If AI is generating most of the code, and engineers are primarily defining intent and validating output, then the relevant knowledge needs to be encoded in specifications and definitions that any team can access.

Instacart is working toward a model where code becomes democratized across groups. Instead of requiring changes to flow through a designated owner team that has deep familiarity with the implementation, the company is embedding domain logic in machine-readable formats that allow multiple teams to request modifications without touching the underlying code directly.

This has implications for how engineering organizations are structured. If code ownership is no longer tied to intimate knowledge of a specific repository, then team boundaries can be drawn around product areas, user flows, or business outcomes rather than technical modules. The risk is that this also makes accountability harder to trace when things go wrong, especially if the AI-generated code introduces subtle bugs that only surface under specific conditions.

The Remaining 3%

The parts of Instacart's codebase that still require human code review fall into predictable categories: compliance, latency-critical paths, and legacy systems with complex interdependencies. These are the areas where the cost of error is high, where regulatory requirements demand auditability, or where performance constraints leave little room for inefficiency.

Kundu's team is gradually refactoring these systems to make them more modular and thus more amenable to machine-driven development. But the process is slow, and it's unclear how far the model can extend. There are workflows where human judgment isn't just a matter of intent validation but of nuanced trade-offs that resist formal specification.

The broader question is whether the 97% figure represents a ceiling or a waypoint. If most new code can be generated and maintained by machines, the engineering profession shifts from a craft centered on building to one centered on orchestration, evaluation, and exception handling. The engineers who thrive in this environment will be those who can clearly articulate constraints, design robust evaluation frameworks, and identify the edge cases where automation breaks down.

What This Means for the Industry

Instacart's approach offers a preview of how engineering work might evolve as generative AI becomes more deeply integrated into development pipelines. The company is not simply using AI as a coding assistant; it has restructured workflows, evaluation processes, and organizational boundaries to assume that machines will handle the bulk of code production.

The model raises practical questions about skill development. If junior engineers spend less time reading and writing code, how do they build the intuition needed to debug complex systems or make architectural decisions later in their careers? If domain knowledge is encoded in specifications rather than held by individual contributors, how does an organization ensure that knowledge remains current and accurate as systems evolve?

At the same time, the efficiency gains are real. Instacart is able to iterate faster on new projects, reduce the burden of technical debt, and detect production issues more reliably. The company has effectively turned software development into a problem of defining intent and validating output, rather than crafting implementations line by line.

For engineering leaders across the industry, the challenge is to determine which parts of this model are transferable. Not every company has Instacart's scale, infrastructure maturity, or appetite for organizational restructuring. But the underlying principle that most repetitive, high-volume development work should be machine work is likely to become more widely accepted as the tooling improves and the success stories accumulate.

The tactical work of engineering is shifting. The question is whether the rest of the industry is ready to follow Instacart's lead.