Instacart AI coding push changes how engineers handle tech debt

Instacart AI coding work has reached the point where CTO Anirban Kundu says engineers usually do not read the code being produced, a shift that reframes tech debt and code review inside the grocery delivery company. Speaking at VB Transform 2026, Kundu said AI agents are taking over repetitive, high-volume software work so engineers can spend more time on intent, judgment and exceptions.

Kundu said Instacart’s builders do not read code in 97% of cases. He described a change in the most basic unit of engineering work: engineers are spending less time writing lines and more time getting AI systems to produce the right outcome.

How is Instacart using AI for coding?

According to Kundu, agents now generate much of Instacart’s code and boilerplate, especially for newer projects where software can be generated or regenerated weekly. He said that cadence has changed the company’s view of maintenance burden: inactive parts can be removed and rebuilt rather than carried forward as long-running technical debt.

“The benefit of that is we don't care about tech debt anymore,” Kundu said, describing code that is dropped when it is no longer active and later rebuilt.

The claim does not cover all systems. Kundu said the remaining 3% includes legacy, compliance-related and latency-sensitive workflows, along with dead, inactive or partly active code that still requires human review. Instacart is trying to reduce those pockets through a project called Atoms, which breaks larger systems into cleaner modules. Kundu said the team began with monoliths and is moving toward architectures driven by remote procedure calls.

The harder problem is evaluation. If AI is producing syntactically correct code at scale, traditional review of individual lines becomes a weaker control point. Kundu said Instacart is pushing engineers toward an intent model, where developers learn how to ask models for the right outcome and rely on separate evaluation systems to test whether the result matches the goal.

Instacart runs about 7,000 automatic evaluations each month, Kundu said. Its system also handles more than 8,000 real-time developer questions with about 99.9% accuracy, according to the company’s figures. Those numbers describe internal performance claims, not an independent benchmark.

What is Instacart’s AI SRE system doing?

Instacart has also built an agentic site reliability engineering system trained on the company’s own incident history and root-cause analyses. An AI SRE system is software that monitors production signals, alerts and engineering discussion, then suggests likely causes or mitigations for outages and degraded service.

Kundu said the company trained the system on the ways Instacart’s own infrastructure has failed, rather than on generic outage examples. He said accuracy in detecting and mitigating production issues rose from about 60% to more than 90% after the system was deployed.

One internal tool, called Blueberry, watches more than 200 Slack channels and looks for patterns across alerts and human discussion, Kundu said. In one incident, a database shard backed by an EBS volume had a temporary problem. Kundu said human responders were not initially focused on AWS disk behavior, but roughly 20 minutes into the incident Blueberry posted in Slack and tied the issue to a blip and an inadequate rollout system called roulette.

Kundu used the example to argue that the AI system can sometimes beat human intuition because it can examine more signals at once. That is a useful production claim, but it is still bounded by the quality of the company’s training data, incident records and escalation process.

What changes for engineers?

Kundu described the engineer’s role as moving toward supervising evaluation, running multiple experiments, handling constraints such as limited top-of-funnel traffic for tests, deciding when to escalate and identifying edge cases.

He also said Instacart is trying to encode domain expertise into definitions and specifications that more teams can use, rather than routing changes through a single owner team. The stated goal is to make code more accessible across groups while keeping the relevant business rules explicit.

This story draws on original reporting from VentureBeat.