Meera Bhatia used to write about 400 lines of code a day. That was her benchmark — roughly two pull requests, most of them small refactors of a payments service she had maintained since joining a Bengaluru fintech in 2021. When we meet her at a café in Koramangala in March, she tells me, almost sheepishly, that she has not opened her editor in six weeks.
"I review diffs now," she says. "I write prompts. I argue with the CTO about which model to put in production. I haven't touched a keyboard the way I used to since Q4 last year." She pauses. "My salary went up 38 percent."
The new job description
If you believe the surveys — and for once, most of the industry seems to — somewhere between 30 and 60 percent of professional software developers now spend the majority of their working hours not writing code, but directing systems that do. A March GitHub report put the figure at 54 percent among developers at companies with more than 500 employees. At smaller Indian startups, where tolerance for experimentation is higher, the number climbs above 70.
What those developers actually do, though, has remained surprisingly difficult to pin down. "AI-first" is a phrase that covers everything from "we use Copilot" to "we have a single junior engineer and fourteen Claude agents on a Kubernetes cluster." We spent a week with five developers at companies that cluster firmly toward the latter end of the spectrum.
"The job isn't gone. It's become a different job. The question is whether the people who trained for the old one will want to do the new one." — Meera Bhatia, Staff Engineer
What changed in 12 months
The single biggest shift, according to every engineer we spoke to, isn't the quality of the AI — it's the quality of the harness. Twelve months ago, an agentic coding system required custom tooling, careful supervision, and a willingness to accept that maybe one in three attempts would produce usable output. By early 2026, the equation had inverted: most startups we surveyed said their agents shipped production-ready pull requests on the first try more than 80 percent of the time.
"The bottleneck is no longer the model," says Raghav Subramanian, a principal engineer at a Mumbai-based B2B SaaS company. "It's the specification. You still need someone who can decompose a problem into something an agent can solve without drifting. That person used to be the tech lead. Now every engineer has to be that person."
The five engineers
The engineers we spent time with were not hand-picked ideological converts. Three of them — Meera, Raghav, and a third, Anuradha Pillai, of an edtech in Hyderabad — had been vocally skeptical of "vibe coding" as recently as last autumn. The other two came in more bullish but had, by their own admission, tempered expectations along the way.
Meera — Payments
Meera's day now starts with a review queue: twelve to eighteen patches generated overnight by a fleet of agents tackling a backlog of low-priority bugs and refactors. Her job is to decide which of them to merge, which to send back with notes, and which to kill. She estimates the decision-making alone consumes four hours of her day.
Raghav — Infrastructure
Raghav designs guardrails. His team's agents have access to production infrastructure, and the most important code he writes is the code that decides what the agents cannot do.
The uncomfortable truth
The uncomfortable truth, as everyone we interviewed acknowledged off the record, is that the new role is not for everyone. It demands a strange combination of skills: systems thinking, editorial judgment, comfort with ambiguity, willingness to defend a decision you didn't fully make. Developers who thrived on the pure craft of writing code — the flow state, the deep focus, the satisfaction of a well-named variable — are, by multiple accounts, the ones quietly leaving the field.
Where this leaves juniors
Entry-level hiring has cratered. Nearly every engineering leader we spoke to admitted they had paused or slowed junior recruitment — not out of callousness, but because the traditional onboarding path (small tasks, tight supervision, gradual increase in scope) no longer maps onto the work. The tasks that used to be the training ground have been automated. What's left requires judgment that is hard to teach except through years of doing.
How the industry solves this — or whether it solves it at all — is, at the moment, the single most-asked question in the rooms where these conversations happen. No one has a good answer. Most people we spoke to suspect the answer will come the way most industry answers do: late, uneven, and at considerable human cost.
Arjun Iyer is a senior correspondent at Citytrends covering technology and labour. Reporting for this piece was done in Bengaluru, Mumbai, and Hyderabad between February and March 2026.
Comments (247)