At VB Transform 2026, Anirban Kundu, Chief Technology Officer of Instacart, posed a provocative idea: what if the bulk of engineering work done today should actually be handled by machines? Kundu argued that developer teams waste time on draining, repetitive, high-volume tasks that AI agents could absorb, freeing humans to concentrate on problems that require judgment, intent and exception handling.
Changes in coding practice
According to Kundu, 97% of Instacart’s builders no longer regularly read code. AI handles most code generation and boilerplate, especially on newer projects where code is generated or regenerated frequently. "In the past, the tactical level was the creation of the code," Kundu said. "In the most tactical level going forward, it's going to be, ‘How do you navigate around the AI system to give you what you want?’"
One practical benefit the company cites is reduced concern over technical debt: inactive parts of the system can be dropped and rebuilt, similar to prior practices of producing assembly or object code.
Why not 100% automation?
About 3% of work still requires human attention. These cases include legacy systems, compliance-sensitive workflows, latency-sensitive components, or large volumes of dead or half-active code. Such situations continue to demand careful human oversight.
Instacart is addressing these areas via a project called Atoms: breaking down monoliths and reconstructing them in cleaner, more modular forms, shifting toward remote procedure call (RPC)-driven architectures.
Evaluations and the shift to an "intent model"
With AI generating code, traditional code reviews become less central — "the lines of code are going to be correct, the syntax is going to meet your expectations," Kundu noted. The new focus is on an "intent model": training developers to ask the right questions of different models from an intent perspective.
Evaluations are performed independently: Instacart runs roughly 7,000 automatic evaluations each month, and its system answers over 8,000 real-time developer queries with about 99.9% accuracy.
Agentic SRE and the Blueberry example
Instacart built an agentic site reliability engineering (SRE) system trained on the company’s own years of incident and root-cause analyses rather than on generic failure data. Instead of teaching a model how outages work in the abstract, the team fed it specifics about how Instacart’s systems have failed and how humans diagnosed and fixed those failures over time.
As a result, the company reports that accuracy in detecting and mitigating production issues rose from roughly 60% to more than 90%.
Kundu cited an example involving Instacart’s internal tool Blueberry. The AI SRE watches roughly 200 Slack channels, monitors signals, and searches for patterns across human conversations and alerts. In one incident, a database shard backed by an EBS volume experienced a hiccup; the human team did not immediately suspect AWS disk issues and were scrambling to find the cause. About 20 minutes in, Blueberry posted on Slack, highlighted a specific blip and tied it to a feature-flag-like system called "roulette" that had been rolling out too aggressively. Blueberry identified the cause and the team resolved the incident. Kundu suggested human intuition might have been slower in that case.
He described a "first brain–second brain" tendency where humans default to familiar patterns and then debug, while the agentic SRE can be more comprehensive in surveying everything and deciding what matters.
Redefining the engineer’s role
Looking forward, the most tactical work for engineers will be navigating AI systems: designing and supervising evaluation processes; coordinating multiple simultaneous experiments and features; managing constraints such as limited top-of-funnel traffic for testing; deciding when to escalate; and identifying edge cases and potential failure points.
Domain expertise is also being rethought. Instead of funneling changes through a single "owner" team that touches the code, Instacart is embedding domain knowledge into definitions and specifications that any team can use. "We’ve lived in this world where this group or this engineering team is the one that can touch the code and make the modification," Kundu said. "We're trying to move into a world where the code becomes completely democratized across groups."
The transition Kundu outlined suggests that widespread deployment of AI agents could reshape engineering work, shifting emphasis from automatable repetitive tasks to human responsibilities such as judgment, system design and oversight.
