Newsletter · 2026-08-22 · Ashutosh Agarwal

Just Go to Claude and Build It - How They Build - Week of August 22, 2026

Startups and venture newsletter for the week of August 22, 2026. Across the week's founder and VC podcasts, tiny teams posted extreme results (a four-person startup booking $1.5 million in a month, a bank CEO telling his sales chief to just build the product in Claude, a snack-box founder out-coding his own engineers) while investors put a hard number on the AI bill: $100,000 of tokens per engineer and dev teams cut by a third.

How They Build

Week of August 22, 2026: Just Go to Claude and Build It

This week a bank CEO told his sales chief to go build a lending product himself, a snack-box founder out-coded his own engineering team, and a four-person startup booked $1.5 million in a single month. The org chart is quietly becoming overhead, and the AI bill is hardening into a per-person budget that finance can trade against headcount.

For a while, "how they build" was a story about software companies. This week it stopped respecting that boundary. The people describing radically smaller, faster ways of running a company included a Stockholm engineering-software founder, the CEO of the fastest-growing digital bank in the United Arab Emirates, and a husband-and-wife team that ships international snacks. What they all described was the same thing from different angles: the number of people it takes to build and run something real is collapsing, the job titles in the middle are dissolving, and the person who used to hand work off to a team now just does the work, with a fleet of AI agents doing the typing.

Two numbers frame the week. One is a startup that closed $1.5 million of recurring revenue in its first month with four people. The other is what a leading venture investor now plans to hand each of his best engineers: $100,000 a year in AI tokens, while cutting the size of the team by a third. Put those two together and you get the real theme of 2026: companies are figuring out the exchange rate between a person, a token, and a shipped result.

The Number: four people, $1.5 million in revenue, one month

The single most striking efficiency stat of the week came from Niklas Lindgren, co-founder and CEO of Endra, on A Product Market Fit Show. Endra makes AI software for the engineers who design the guts of buildings: the heating, cooling, electrical and plumbing systems (the industry calls it "MEP," for mechanical, electrical and plumbing). Think of it as the equivalent of what Harvey and Legora are doing for law, but for the people who design how a hospital or a data center actually works.

Lindgren's team was tiny. "Being just a four-man team, none of us being the actual [ideal customer] ourselves, we needed to get out and speak with users," he said, describing how they cold-DMed engineers on Reddit in early 2025 to learn the job. They launched, and then this happened:

"We started our first pilots early January '26 that very fast converted into signed contracts end of January '26… And we closed a million and a half in general [ARR] alone. And that was all enterprise revenue."

That is $1.5 million of annual recurring revenue booked in a single month, from roughly four or five enterprise customers at an average contract value "around 300-something" thousand dollars. The company has since raised a $4 million pre-seed, a $20 million seed, and a $50 million Series A from Andreessen Horowitz in June, a round about three times the size of a typical Series A, which Lindgren noted is normally around $15 million.

What makes the revenue possible is a productivity jump that sounds made up until you hear the mechanics. Building engineers today work in a legacy program called Revit, built in 1997, plus four or five other disconnected tools; a huge share of their day is copying data between programs by hand. For a hospital, engineers manually place "each and every receptacle, each and every lighting fixture, each and every lighting switch… tens of thousands of units," then hand-calculate the electrical loads across spreadsheets and separate software. Endra runs agents that generate the design instead. The moment the pull became undeniable was when customers fed in projects they'd already finished:

"I spent 400 hours doing this project. Endra created this for me in under 10 minutes."

Lindgren was careful about the catch that usually kills these stories: the "80% good enough to be useless" problem, where AI gets you most of the way but the review and rework eat all the time you saved. In Endra's case, he argued, "the 80% to 90% really was 80% to 90%," because building design is iterative by nature: an 80%-complete first pass is a real, billable deliverable, not a rough draft you have to redo. This is also a founder who has done lean before: he and his childhood friend and co-founder Anton bootstrapped a physical security-systems business to roughly $15–20 million in revenue and sold it to private equity in 2022, when they were about 26.

A Product Market Fit Show, "He spent $150K on brand before he had a product, then closed $1.5M ARR in 1 month. | Niklas Lindgren, Co-Founder & CEO of Endra" (August 17, 2026).

What Founders Changed

"Just go to Claude and build it." A bank CEO, to his sales chief. The most quotable version of the everyone-builds shift came from an unlikely place: a regulated bank. On with Loulou, Jayesh Patel, CEO of Wio Bank in the UAE, described how his company now ships products. He'd been in months of meetings about a new auto-loans product and was sick of the back-and-forth. So he turned to his most senior business executive:

"My chief commercial officer, I told him, listen, man, you know what to do. Just go to Claude and build it. And literally, he went to Claude. He built the whole journey."

Wio has internal "AI toolkits" that take whatever someone builds in Claude and standardize it into the bank's real design, data and code systems. It wasn't a one-off. "We have a great lawyer who built an entire SME account opening platform," Patel said. "When we sat through the presentation, we were shocked. We were speechless… this is much better than what we have." They then put four or five people around the lawyer's prototype to ship it for real, with the lawyer reviewing it every few weeks. Patel's conclusion about what this does to the org chart is blunt: "I can be a lawyer and turn into a product builder and I can run the product… suddenly they realize they don't work for a department. They work for a company… Everyone works for Wio, not departments. And I think that's the old corporate culture which is going to be broken."

He was refreshingly clear-eyed about where it lands hardest, the same layer everyone keeps naming. "For the guys in middle management, it is the hardest," he said, because the new hires "have the skill set" natively and the existing managers "think differently." And he pushed back on the doom framing: "Tomorrow's companies are going to be smaller, but there are going to be more companies too."

with Loulou, "Jayesh Patel (Wio Bank): The Playbook Behind the UAE's Fastest-Growing Digital Bank" (August 18, 2026).

A snack-box founder is now out-coding his own engineering team. On the eComFuel Podcast, Eli Zahner, co-founder of Universal Yums (a subscription business that ships snack boxes from a different country each month) gave the most vivid picture of a non-technical CEO becoming the most productive engineer in his own company. Zahner now spends about 90% of his computer time in the terminal and "probably 70% of my time right now writing code," using Claude Code. The kicker:

"I'm currently as an individual pushing more code than my entire engineering team combined."

That team is three full-time engineers and two part-time ones. Zahner's edge isn't that he's a better programmer; it's that he has all the business context and, as CEO, the authority to grab anyone's time on the spot. He's collapsed his whole development process into two reusable instruction files ("a scoping skill and a building skill") so that "the amount of time that it takes me to develop an average feature is maybe 10 or 15 minutes of work." He tried pushing the tool company-wide, with mixed results he described memorably: giving an untrained employee Claude Code is "like you're giving a civilian an AK-47." The curious ones "turned into really super, super users"; others floundered without hand-holding. His fix is a shared "company brain" (they call theirs "yums brain"), a database of the company's knowledge, brand guidelines and skills that both people and background agents can pull from. His framing to his team is the sentence a lot of operators are quietly saying: "Everyone in our organization needs to get either really good with people or really good with AI. There's not a lot of middle ground."

eComFuel Podcast, "An Unreasonably Deep Dive on Building a Claude Code First Company" (August 21, 2026).

"We were just buying revenue with labor. And that was cheating." The most direct attack on the org chart itself came from Luke Girgis, an operator and author of a book literally titled Death of the Org Chart, on 21st Century Entrepreneurship. His opening line sets the tone: "We will look at org charts as an ancient artifact of companies from the past." Girgis took over an e-commerce food business, Provador, that was losing $400,000 a month and heading for bankruptcy. Knowing nothing about the industry, he broke every job down into micro-tasks, found the places where "three people were doing the job of one," and rebuilt the company around workflows instead of headcount. The result: "It did mean I let go of 70% of the staff, but it also meant that the business survived… and became break even." He credits AI, which arrived right as he was doing it, for a lot of the savings: "I don't know that the business would be break even today without that technology."

His most useful insight came from a company he'd run before, publishing Rolling Stone and Variety in Australia, where the margin never got above 4% no matter how fast revenue grew:

"We always had to hire more people for every dollar that came in. So we weren't actually scaling. We were just buying revenue with labor. And that was cheating. That was fake. That was not a real sustainable growth business."

His rule now: never build a company whose org chart has to grow in lockstep with revenue. At his artist-management firm, he says he automated 90% of the administrative work that used to eat half of each manager's day (chasing contracts, moving data between spreadsheets) and "every artist on our roster, bar no exception, is now making more money than they've ever made." His warning to founders bolting AI onto an unchanged org: that's "like driving a Ferrari in traffic… it looks cool, but it does nothing." You have to rebuild the workflow first, then arrange people around it, "rather than have the workflows come from the staff."

21st Century Entrepreneurship, "#541 Luke Girgis: How Do You Turn $400K a Month Into Breakeven?" (August 19, 2026).

A $49 billion company deleted its general managers, and let AI hold the org's memory. It isn't only startups. On The Product Podcast, Willem Avé, Global Head of Product at Square (part of Block, valued around $49 billion), explained why the company tore up its business-unit structure about two years ago and reorganized entirely into functional teams (one big org each for product, design and engineering), killing the general-manager model. The reason was craft: the old structure "reduced the kind of craft and excellence that we wanted to see within each of our functions." But the AI twist is the interesting part. Historically, companies encode hard-won lessons into process and layers of people. Avé's argument is that you no longer have to:

"AI gives us the ability to encode that knowledge at another layer. It doesn't have to be encoded in the org entirely. It can be encoded in data that agents have access to. And that really democratizes decision-making."

The payoff, he said, is that "the best teams are kind of small, fast moving, fully autonomous," able to get an answer from an agent instead of "playing telephone" up and down a hierarchy. He paired it with a "DRI" (directly-responsible-individual) model to cut through the "silent veto." Notably, Avé didn't claim management disappears (see The Other Side below).

The Product Podcast, "Square Global Head of Product on: How to Build AI Agents People Actually Use, Why $49B Company Got Rid of General Managers | Willem Avé" (August 19, 2026).

The Cost Corner: $100,000 in tokens, and cut the team by a third

The theme worth flagging loudest again this week is the money, and it got a genuinely new, concrete framing. On The Twenty Minute VC, investors Jason Lemkin and Rory O'Driscoll did the math out loud on what an AI-era engineering budget actually looks like, and Lemkin put a hard number on it:

"I think we'll give each of our best engineers $100,000 of tokens. And in return, we'll cut the size of our dev teams 30%, 40%."

The mental model: a fully-loaded engineer costs roughly $200,000 in wages and benefits, so you add about $100,000 of AI on top ("for running inference 24-7 with 10 agents in parallel") and in exchange you employ meaningfully fewer of them. That is the exchange rate between a person and a token, stated plainly. And it's arriving fast: "The last 60 days, every single scale-up is capping their AI budget for real. It's not just [the] Ubers of the world. Everyone's capping it… It's $6 million a year, $8 million a year." For a reality check on the spread, they cited spending data from the corporate-card company Ramp: even the top 1% of the most AI-forward companies are spending only about $7,000 a month per person, roughly $0.50 of AI for every $1 of salary at the very pointy end, with an enormous gap down to everyone else.

Two things make this different from last year's "look how big the bill is" panic. First, the output is finally real. Lemkin said his fastest-growing companies "literally are shipping two to three times faster… only recently," and in two recent board meetings the teams had already "finished the roadmap for the year" and were "well into the 2027 roadmap." His provocation: "If you're not into your 2027 roadmap deep into it by August of 2026 in the agentic world, your team is not good enough to survive today." Second, this now maps to a very large, very specific labor pool. There are roughly 1.8 million people writing code in the US and about 5 million doing software-related work, grossing around $600 billion in wages a year, which is why the investors landed on a figure like $200 billion for what AI could pull out of software labor, i.e. "replacing about a third of them."

The supplier side of that equation had a landmark week too. Anthropic turned its first profit, on $11.5 billion of second-quarter revenue: a company that did $4.5 billion in all of last year is now doing roughly $10 billion in a quarter. As the hosts put it, when gross margin swings from negative to around 40% while revenue grows more than tenfold, "you can't add expenses below the line fast enough to stop yourself making money." (Fine print worth remembering: that profit sits alongside enormous off-balance-sheet commitments to buy future compute, and stock-based pay "like we've never seen." Anthropic's own plan reportedly targets $200 billion of revenue in 2028 and $600 billion the year after, numbers that only work if the world really does hand AI a huge slice of every knowledge worker's paycheck.)

And the clearest sign of how much the token economics still hurt: SpaceX closed a $60 billion, all-stock takeover of the coding tool Cursor. For much of its life Cursor was the classic money-losing AI business, "selling a dollar's worth of tokens for 80 cents." What made it a fit for Elon Musk specifically was that he owns a giant, underused pile of computing power (his "Colossus" cluster). As one host framed the buyer's logic: "Your gross margin problem is my revenue opportunity." In other words, the thing that made Cursor a bad standalone business (spending more on compute than it collects) becomes an asset when the owner is also the one selling the compute. That is the whole cost story of 2026 in one sentence: the AI bill is somebody else's revenue, and everyone is now negotiating over who eats it.

The Twenty Minute VC, "SpaceX Buys Cursor for $60BN | Stripe's $8BN OpenRouter Bet | Anthropic's First Profit & The Math Behind Reaching $600BN in Revenue? | Lovable and Higgsfield Raise Mega Rounds" (August 20, 2026).

The Other Side: smaller isn't the same as gutted

For balance, three of this week's own sources pushed back on the "cut everyone" reflex, not by denying the leverage, but by complicating it.

Flatter doesn't mean no managers. Even as he praised small, autonomous teams, Square's Willem Avé was explicit that the shift "doesn't diminish the need of management, right? People grow in their careers… you need to also have incredible managers (people managers) that can help people through their kind of career growth." His model marries decisive individual owners with real managers, not one instead of the other. (The Product Podcast, August 19, 2026.)

Fewer people per company, but more companies. Wio's Jayesh Patel made the case that lean orgs don't have to mean mass unemployment: as AI frees up capable people, "they go after everything else that they haven't pursued," and markets that looked fixed "become two, three X." His frame is multiplication, not just subtraction: smaller companies, but a lot more of them, and existing stars stretched across more of the work rather than replaced. (with Loulou, August 18, 2026.)

And buying the tool isn't the same as getting the value. On DataFramed, Jennifer Smith, CEO of Scribe, warned about the gap between an AI purchase and an actual result. Her example: a team bought an AI ad-buying tool, wrote the CFO a glowing return-on-investment memo ("it's going to make us more accurate and faster… our team can handle more, it looks really good on paper"), and then "nobody changed any of their behavior and we just spend money on this tool for no reason." Her fix is to measure whether the workflow actually changed and whether it produced the promised result, not to celebrate adoption. It's the quiet counterweight to every productivity-multiplier claim in this issue: more capability only counts if the way people work actually changes. (DataFramed, "#373 What Do Your Colleagues Do All Day?… | Jennifer Smith, CEO at Scribe," August 17, 2026.)

The through-line: The unit of a company is shrinking. A four-person startup can book seven figures in a month; a bank's lawyer can build a product on a weekend; a snack-box CEO can out-code his own engineers. As that happens, the org chart (the layers, the handoffs, the "who reports to whom") starts to look less like structure and more like cost, which is exactly why an operator this week could publish a book called Death of the Org Chart and have it ring true. But the savings don't vanish into thin air. They reappear on the token meter, where a single engineer's AI habit now runs at half a second salary, and where the sharpest question a founder can ask has quietly changed. Last year it was "how do we use more AI?" This year it's the one every CFO is now asking out loud: for the people and projects still here, what's a token worth, and who's actually earned theirs?