Instacart is posing the provocative query: What if many of the work your engineers do at this time ought to, in actual fact, be accomplished by machines?
At VB Rework 2026, CTO Anirban Kundu argued that dev groups proceed to waste their time on draining, repetitive, high-volume work; this needs to be absorbed by AI brokers in order that people can concentrate on issues that require judgment, intent, and exception dealing with.
In truth, in 97% of instances, Instacart’s builders don’t even learn code anymore.
“Up to now, the tactical degree was the creation of the code,” Kundu stated. “In essentially the most tactical degree going ahead, it’ll be, ‘How do you navigate across the AI system to present you what you need?’”
AI producing code, performing “fairly severe evals”
That doesn’t imply people by no means have a look at code; brokers deal with the majority of code technology and boilerplate, notably with newer tasks the place code is generated or regenerated on a weekly foundation.
“The advantage of that’s we do not care about tech debt anymore,” Kundu stated. “Issues that aren’t lively simply get dropped out after which it will get rebuilt, sort of like how we used to construct meeting code or object code.”
So why not 100%? The remaining 3% is in legacy, compliance, and latency-sensitive methods and workflows, or pushed by a “boatload of code” that’s lifeless, not lively, or half-active. These instances nonetheless want cautious human consideration.
Instacart is slowly “smoothing these components out,” nonetheless, breaking methods down in an aptly-named challenge Atoms, then constructing them again up in a cleaner, extra modular kind. Kundu’s staff began with the “monoliths” and is shifting to distant process name (RPC)-driven architectures.
However analysis stays one of many overarching challenges. Code critiques aren’t as related when AI is producing code — as Kundu famous, “the traces of code are going to be appropriate, the syntax goes to satisfy your expectations” — so the aim is to maneuver to an “intent mannequin.” That’s, coaching devs to allow them to ask totally different fashions the correct questions from an intent perspective.
Evals are then carried out independently: Roughly 7,000 automated evaluations run every month, and the system solutions 8,000-plus real-time developer queries with about 99.9% accuracy.
Figuring out “hiccups” that human instinct might need missed
Dovetailing with this, Instacart has constructed an agentic website reliability engineering (SRE) system skilled on years of the corporate’s personal incidents and root-cause analyses slightly than generic failure knowledge. As a substitute of educating a mannequin how manufacturing outages work within the summary, the staff fed it the precise methods Instacart’s methods have damaged over time, together with the methods people identified and stuck them.
Because of this, the corporate has seen accuracy in detecting and mitigating manufacturing points bounce from roughly 60 to greater than 90%.
Kundu pointed to at least one instance with Instacart’s inside device Blueberry. The AI SRE colleague watches 200-some-odd Slack channels, screens indicators, and appears for patterns throughout human conversations and alerts.
In a single incident, a database shard backed by an EBS quantity that had a “hiccup” for a time frame. The human staff didn’t instantly suspect AWS disk points and have been “clearly scrambling” to determine why this specific shard misbehaved.
However about 20 minutes in, Blueberry posted on Slack, pointing to a particular blip and tying it to a feature-flag-like system known as “roulette” that had been insufficient. “It is presupposed to be rolling out on this cadence, [but] it had been an excessive amount of,” Kundu stated.
Blueberry figured it out, and the staff resolved the incident. “Would have a human been as fast? I believe the issue is human instinct would maintain us again somewhat bit,” Kundu stated.
People are likely to default to patterns we’ve seen earlier than, then resort to debugging; Kundu known as this the “first brain-second mind sort of factor.” However Instacart’s agentic SRE is definitely “extra complete in its capacity to have a look at every thing after which be capable to determine what does or would not matter.”
Redefining the engineer’s job
Trying forward, essentially the most tactical work for engineers will probably be navigating AI methods: Designing and supervising analysis processes; coordinating a number of simultaneous experiments and options; managing constraints like restricted top-of-funnel site visitors for testing; determining when to escalate; figuring out edge instances and the place issues may break.
Area experience can be being rethought within the age of AI. As a substitute of bottlenecking modifications via a single “proprietor” staff that touches the code, Instacart is embedding area data into definitions and specs that any staff can use.
“We’ve lived on this world the place this group or this engineering staff is the one that may contact the code and make the modification,” stated Kundu. “We’re making an attempt to maneuver right into a world the place the code turns into utterly democratized throughout teams.”
