Myntra Founder: Is This the End of Software Engineers?

ELI5/TLDR

Two old colleagues — Mukesh Bansal (Myntra, Cure.fit) and Piyush Ranjan (ex-Flipkart CTO, ex-Google VP) — compare notes on what AI has done to building software in the last six months. Both are CEOs who have stopped delegating and started writing code again, because they can. Their case: an engineer is no longer the bottleneck, distribution is; a 10,000-person knowledge-work company no longer makes sense; the new managerial skill is managing AI agents. The conversation is part exuberance, part working notes from inside two companies running the experiment.

The Full Story

Two CEOs who started coding again

Piyush built Enrico, an AI chief of staff at his education startup Furmi, during a two-week cruise. Mukesh, who hadn’t written code in twenty years, watched him do it and ignored the pull for two months. Then he gave in. A month later he had built the V1 of a new product his team had failed to ship for $100,000 with an outsourced agency the previous year. His own version cost about $5,000 in API credits.

“I think I’ll build a better product than that in less than 10 days. About 4, 5,000 dollars of credit burned.”

The thing both keep coming back to is that this isn’t a tools upgrade. It’s a category change. Piyush calls it the collapse of the gap between imagination and instantiation. Mukesh frames it as a workflow death sentence for everyone who used to lead with PowerPoint:

“If you have a compelling idea, just build the whole damn product in a week. It wouldn’t be surprising when it is in a day.”

The line that lands hardest in the episode is Piyush’s: “We are no longer in the world we were 3 months ago.” It’s also the line that’s hardest to verify. Most of the evidence here is anecdotal — Enrico, Mukesh’s V1, a New York GLP startup with two engineers and $1.8 billion in sales. But the anecdotes are stacking up.

Revenue per employee, redrawn

The metric they keep using is revenue per employee. Google and Facebook used to brag about $1 million per employee. Vertical AI-native startups, Piyush claims, are landing at $3.5 million to $5 million. Furmi, with about ten people, says it gets the output of a five-times-larger team.

Then they turn it on the Indian IT services industry. TCS: 700,000-odd employees, ~$30 billion revenue, roughly $40,000 per employee. The implication, unsaid but obvious: the business model assumes a price gap between a Bangalore engineer and a US engineer. AI doesn’t care which side of that gap you’re on. It eats both.

“If it is not done by the incumbents, it’ll be done by somebody else with disruption.”

Piyush’s optimism on the displacement question is genuine but theoretical: of 70,000 displaced, maybe 7,000 start companies. The 63,000 are not addressed. The conversation moves on quickly, which is its own kind of answer.

The death of the org chart

The argument here is unfussy. Hierarchies exist for information flow. If an AI like Enrico holds all company information and is accessible to anyone, the hierarchy’s job is gone. Furmi has reorganized around what they call “problem-solving loops” — a problem, the people motivated to solve it, the AI tools, no boxes.

Piyush goes further: “There are no more managers. Everybody is an IC.” Decision-making in ambiguity still needs someone — the Amazon “single-threaded leader” pattern — but that doesn’t require headcount underneath them. Mukesh, who has run companies for two decades, agrees and says he wants to push his own manager-time below 20% within months.

This is the most provocative claim in the episode and the one most likely to age badly or age beautifully. Either nobody manages anyone in five years, or the people writing the code in 2026 quietly rebuild a thinner version of the org chart by 2028 because coordination is hard.

What you hire for now

If the org chart is dead, what does the JD look like? Piyush’s answer is two-part:

Lived experience — not aggregable expertise, but the specific texture of having done the thing before. The kind of knowledge that doesn’t show up in LLM training data.
Generalist by attitude, not qualification — the engineer who’ll do UX, the designer who’ll write SQL. Not a job description; a disposition.

Mukesh adds a third filter: flexibility. He notes that Furmi and Mirabilis now make candidates do work trials, which has become an inadvertent senior-people-screen. The senior candidates who say “I don’t interview that way anymore” self-select out, which is exactly the signal he wants.

The deeper claim sitting underneath all of this is that “intelligence is utility.” Smartness is now something you rent. So interviews shouldn’t test it.

“When intelligence is utility, what is really at premium is the question.”

Curiosity and agency, repeated like a refrain. The person willing to ask and the person willing to act. That’s the hire.

The board of advisors made of LLMs

The most concrete tool Piyush describes is a virtual board of advisors. He had Enrico instantiate five personas — Steve Jobs, Peter Thiel, Jeff Bezos, Charlie Munger, Sam Altman — running on Gemini 3.1 Pro because Google had ingested the most multimodal data on these people (interviews, books, videos).

The mechanic that makes it not a parlour trick: Enrico continuously distills Furmi’s meeting notes into a strategic memory. When Piyush convenes the board, that context goes in as a pre-read. The advisors then ask him questions, push back on each other, force him to commit to positions.

“It’s not a fire and forget. It’s a board of advisor meeting. So they will look at it and then they will ask question. Hey Piyush, so what do you think?”

Mukesh, mostly skeptical of fancy AI demos, presses on whether this is gimmicky. Piyush’s defence is that the loop is the value, not the personas — the structured pre-read, the multi-perspective debate, the forced commitment. The personas are a thinking aid. He used the output for actual pricing decisions on Furmi’s pilots and tested it against his leadership team. They couldn’t poke holes.

This is the part of the episode worth lingering on, because it’s the only concrete workflow innovation described in any detail. Most of the rest is “AI is fast and good.” This is “here is an actual structure.”

Memory, agents, the things that don’t work yet

Both of them are honest about what’s broken. Memory is the headline problem. Mukesh says his chief of staff keeps forgetting things he has explicitly told it. Piyush has a poetic frame for why memory matters: consciousness, in some neuroscience theories, is global memory across neurons. Give an AI persistent shared memory across a team and you’ve raised its level of operation.

Funnily, the solution converging across the field is not exotic — it’s flat files. The simple thing turning out to be the right thing.

Two other open frontiers they flag:

Long-horizon agents that stay on task for a day instead of drifting after twenty minutes
Embodied AI — the blueberry-sorting problem. Robots that can hold a blueberry without crushing it. Sensing and actuating, not planning, are now the bottleneck.

Innovation as commodity, distribution as moat

The economics shift Piyush flags is one Shantum will already have thought through but it’s worth saying cleanly: software’s marginal cost is no longer zero. API calls cost money per request. So the economics of “build once, distribute infinitely free” is gone. You have to design for cost from day one.

What this does to startup strategy: if anyone can build anything cheaply, then innovation is no longer the moat. Distribution is. Big companies have it; startups don’t. So the conventional wisdom flips — distribution becomes precious.

“If innovation becomes a commodity, then distribution becomes a precious thing.”

But Piyush has a contrarian read on the implication. Big tech’s distribution doesn’t automatically win because AI is personalized. If the chatbot you’ve trained on yourself for two years fits you like a glove, you won’t switch even for a better product. Switching cost goes up the longer you use a tool. So the future may look more like a thousand small AI-native firms — “10 engineers with a billion-dollar corner” — than three trillion-dollar giants. He compares it to the law-firm model: large by aggregation, not by single-asset leverage.

Mukesh, an avowed big-tech skeptic, likes the vision but pushes back. People still want what other people have. Hollywood still wins. Top 100 books still capture the bulk of reading hours. Brand and trust are sticky, and incumbents will leverage them. His counter-counter: incumbents move slowly and can’t unlearn. So maybe they fumble the moment.

The cognitive debt question

The most measured piece of the conversation is on the cost of using AI well. Both grant that AI use can produce intellectual atrophy. Piyush cites an MIT paper showing students using ChatGPT see grades and test scores diverge — implying surface performance up, real capability down. Furmi’s product thesis is built around this: how do you preserve the productive struggle of learning while still using AI?

Mukesh’s translation of the risk is sharper:

“You become intellectually lazy and instead of you commanding AI, AI now commands you.”

The two skills they argue protect against this: critical thinking (be constructive and skeptical at once) and systems thinking (zoom out, see second- and third-order effects, model the whole). Both of which the education system, especially higher education, actively trains out of people by pushing them deeper into specialization.

The 12-month picture

Mukesh closes with a forecast question. Piyush’s bet is that planning and architecture-level thinking will get tackled, and embodied AI — robotics — gets to within arm’s reach within twelve months. Not everywhere, but visibly arriving. He cites the William Gibson line about the future being unevenly distributed.

The other thing both want to be done with is meetings. Piyush has compressed his manager time to one hour a day. Mukesh wants to get there. The fantasy is that meetings die because the AI does the coordination. Whether that’s true or just what two engineers-at-heart want to be true is left to the listener.

“I’ve always felt that meetings are the curse of corporate life. I think we are on the verge of being delivered from that curse.”

Key Takeaways

Revenue per employee in AI-native startups: $3.5M–$5M, vs. $1M for classic tech, vs. ~$40K for Indian IT services. The gap predicts industry compression.
Hire for lived experience and generalist disposition. Stop testing for retrievable expertise — that’s now a utility.
Curiosity + agency is the actual job spec. Expertise without those two is dead weight.
“Problem-solving loops,” not org charts. Furmi’s structure has no managers; people self-organize around problems with AI tools available.
The non-technical / technical founder distinction is dissolving. Anyone can build the V1.
Innovation is becoming a commodity; distribution becomes the moat. But personalization may erode incumbents’ distribution advantage by locking users into AI tools that fit them like a glove.
Software now has non-zero marginal cost. Architect for unit economics from day one.
Memory is the unsolved bottleneck for AI agents. Long-horizon task focus and embodied dexterity (the blueberry-sorting problem) are the next frontiers.
Use the most expensive model where mistakes are costly (Piyush uses Opus 4.6 for coding); use cheap or open-source models for routine cron jobs.
Build a “virtual board of advisors” using a model with deep multimodal context (Gemini 3.1 Pro) — feed it distilled meeting memory, let personas debate. Not a parlour trick if the pre-read is real.
Cognitive debt is real. AI users get measurably worse at independent thinking. Critical thinking and systems thinking are the two skills that protect against it.
Indian IT services has the largest displacement risk — TCS-style $40K-per-employee model can’t survive vertical AI-native competitors at $3.5M per employee.
Open-source models (especially Chinese: DeepSeek, MiniMax) are within 10x cost and near-parity for many tasks.

Claude’s Take

Two CEOs who admire each other talking about how good AI is. There is no skeptic in the room. That’s the structural limitation of the episode, and you should price it in.

That said, the value here is concentrated in three places. First, the org-design observations — problem-solving loops instead of hierarchy, no more managers, work trials as a senior-candidate filter — are concrete enough to test. Second, the virtual board of advisors workflow is the only described AI tool in the conversation that goes beyond “I asked Claude and it built it.” It has a real architecture: persistent distilled memory, multi-perspective debate, forced commitment. Worth borrowing. Third, the inversion of the innovation/distribution moat is a clean way to frame what’s actually happening to startup strategy.

Where to be careful. The “world has changed in 3 months” framing is the kind of thing CEOs say when they’re inside a hype cycle. The revenue-per-employee numbers are cherry-picked from the absolute frontier of vertical AI startups, not a representative sample. The displacement-becomes-entrepreneurship argument (70,000 → 7,000 startups) is an optimistic shrug, not analysis. And the “no more managers” claim fights against everything we know about how organizations of even ten people actually function — Furmi may genuinely have none, but at fifty people, someone reinvents coordination.

Piyush is more interesting than Mukesh in this conversation. Mukesh asks good questions but also occasionally rounds them off into agreement before they bite. The Hollywood-and-top-100-books pushback on Piyush’s anti-incumbent thesis is the rare moment where the conversation has actual friction, and you can feel both of them sharpen.

The one piece of practical advice that survives the hype tax: stop pitching, start building. If you have an idea, prototype it before you write a deck. The marginal cost of trying is now too low to gate it with planning. That’s not novel but it lands harder when said by two people running real companies, with examples.

Score: 7. Solid signal-to-noise from two operators with skin in the game, weakened by the mutual-admiration format and the hand-wave on displacement. Worth listening for the org-design and board-of-advisors bits; skim the rest.

Myntra Founder: Is This the End of Software Engineers? | SparX