Newsletter · 2026-08-29 · Ashutosh Agarwal

The Engineer's New Job Is Keeping the Agents Busy While You Sleep - Week of August 29, 2026

Startups and venture for the week ending August 29, 2026. Podcast synthesis on Robinhood's CEO describing engineers who now manage fleets of coding agents around the clock, a Microsoft-backed founder cutting a dashboard build from ten hours to ten minutes while warning that gutting the junior bench is short-sighted, an operator building an org chart out of digital employees, and the running meter that breaks per-seat pricing.

Startups & Venture

Week of August 29, 2026: The Engineer's New Job Is Keeping the Agents Busy While You Sleep

This week the job of a startup engineer quietly changed. You don't write the code anymore, you run the robots that do, and the smart ones never clock off.

The Number: "Close to 100%"

That's how Robinhood CEO Vlad Tenev describes AI adoption inside his own company, a publicly traded brokerage with millions of customers and real money on the line, not a five-person startup in a garage.

Here's the part worth screenshotting. Asked how much AI has changed the company, Tenev didn't talk about a chatbot or a productivity tool. He described the engineering job itself being rewritten:

The workflow has shifted from humans writing the code and humans reviewing the code to generally humans overseeing agents. And those agents are writing the code and reviewing the code. And your job as an engineer is shifting to making sure your agents are well managed and seeing how many agents you can employ and can you keep them busy all the time, including when you're sleeping.

Read that last line again. The new performance metric isn't lines of code or pull requests. It's how many tireless software agents one human can keep productively occupied, around the clock, including overnight. The best engineers, in this telling, aren't the fastest typists. They're the best managers of a workforce that doesn't sleep.

The Iced Coffee Hour, "Robinhood CEO's Shocking Prediction On Stocks, The AI Bubble, & Gambling Controversy | Vlad Tenev" (Aug 23, 2026).

To Tenev's credit, he refused the lazy conclusion that this means firing everyone. Asked whether a handful of people could soon run the whole company from home, he pushed back: it's "not our goal," "someone still has to be running the agents," and Robinhood has been "focused on doing more with the teams that we have and in giving people more leverage" rather than "just reducing headcount." That's the honest version of this whole story: the same tool that lets one person do the work of ten is also, right now, being used to do more, not to shrink the payroll. Which of those two things wins is the real question hanging over every founder this year.

What Founders Changed

A whole company built on a "toolbox," not a product, and non-engineers doing the building. Husein Sharaf founded Cloudforce, a Microsoft-backed firm (grown "from a one-person startup") that builds AI systems for hospitals, universities and government. Instead of shipping one narrow tool, he built a platform his customers assemble themselves, letting them "grab Legos and put the Legos together." The concrete results he cited are the kind of numbers that make a CFO sit up:

• A finance staffer who used to build a Power BI dashboard by hand now asks the platform to write the underlying code for them and gets it done "in 10 minutes versus 10 hours."
• A university solved an admissions bottleneck with "three agents and orchestrator on top" that "reduces the time it took from three days to 30 minutes."

Sharaf also runs an internal ritual he calls "Frontier Fridays," where every week the whole team shares one mundane way they used AI that week, deliberately without trying to measure the dollar return. His argument: nobody asks "what's the ROI of a Google search" anymore, and AI is heading for the same status, a baseline expectation rather than a line item you justify. The AI Files, "A conversation with Husein Sharaf, Founder and CEO of Cloudforce" (Aug 23, 2026).

An operator building an actual org chart out of AI agents. On a show for entrepreneurs, coach and operator Peter Swain walked through how he's replacing job functions in his own business with what he calls "digital employees," each given "a role, a goal, and a backstory," and (for his own amusement) named after Mortal Kombat characters. The standout example: an agent called Raiden whose only job is to chase down leads who applied to his program but never booked a call.

This job is worth north of $200,000 a year to me.

Swain isn't stopping at one. He described a future with "400 digital employees," too many to manage directly, so he's planning a management layer of agents: a digital "VP of sales" that supervises the junior agents and reports up, so "I'll probably be talking to one agent in charge of sales." It's an org chart where most of the boxes are software. Reality check: Swain is a solo operator and coach, not a scaled startup, and he was candid that setting up the first agent still takes real work (his class put 75 people through it in a weekend, and all 75 walked out with a working agent, but it took a weekend). Still, the mental model, hiring, naming, and managing agents like staff, complete with middle managers, is where a lot of small businesses are visibly heading. Successful Idiots | AI for Entrepreneurs, "The Difference Between an AI Agent and a Digital Employee" (Aug 23, 2026).

Even the book got vibe-edited. Silicon Valley veteran Keith Teare described using an AI agent to cut his book from 80,000 words to 60,000: the agent built him a two-panel editing view, and he went through "70 chapters... about five-ish chapters a day," finishing in three days. He says he wrote the original draft "in about two days." A small anecdote, but it rhymes with the week's theme, where the human sets the goal and steers and the machine does the grinding. Keen On America, "We Are (Not) the Champions: Why Silicon Valley Isn't Providing an AI Human Dividend" (Aug 23, 2026).

The Other Side

For balance, three people this week pumped the brakes, and they made the sharpest arguments of the bunch.

"It's really short-sighted." The same Cloudforce founder who loves what AI does to productivity had a blunt warning about the fashionable move of cutting junior staff and replacing them with agents:

A lot of organizations are cutting headcount and replacing certain low-level roles with AI. And I think that it's really short-sighted, especially because if all you have is senior people, eventually those senior people age out. And then what are you replacing with? Like you need junior people to become the mid-level people to become the senior people.

His point: if the senior folks are becoming "orchestrators of agents," effectively "the HR department for AI," someone still has to grow into those senior seats. Cloudforce, he said, is "hiring more rapidly in this moment than we ever have," juniors very much included, on the theory that a junior who's fluent with AI is "so much more valuable so much more quickly" than a junior was three or five years ago. Cut the bottom of the pyramid now, and you've eaten your own future. The AI Files (Aug 23, 2026).

"SaaSpocalypse is dramatically overplayed." Jeff McQueen, a serial founder who sold his last software company to private equity and now runs Worksites AI, took direct aim at the popular idea that cheap AI coding will kill packaged software because everyone will just build their own. His analogy was gardening: yes, YouTube and AI can teach anyone to garden, but "10 or 20% of us... love this stuff and do it for free... the rest of the world wants the solution. They want the outcome. And they do not have any interest whatsoever with learning shell commands to run up NPM projects." Most companies, he argued, will keep buying software, not roll their own. The AI for Sales Podcast, "The Future of AI in Sales and Business" (Aug 22, 2026).

The founder's new job: babysitting the robots. Keith Teare put words to a feeling a lot of builders are having, that running agents isn't effortless leverage but a new kind of constant work. Pointing to reporting that founders are "working harder than ever to keep up with their agents," he described his own days: "You can't run AI agents without engaging with them pretty much constantly... my productivity has gone through the roof, but my work with agents has replaced what I used to do. It's like being Captain Kirk on the Starship Enterprise... flying the thing." The leverage is real. So is the treadmill. Keen On America (Aug 23, 2026).

The Cost Corner

The uncomfortable math under all of this got aired out again this week: the software you build with AI has a running meter attached, and the old pricing playbook wasn't designed for it.

The per-seat SaaS model is quietly broken. Cloudforce's Husein Sharaf made the clearest version of the argument:

The traditional software model doesn't really work in the age of AI at all... this per user per month license was never designed to really consider the cost of inference. And the cost of inference is expensive.

His twist on the "but models keep getting cheaper" comfort blanket: cheaper models don't save you, because you use more. "As GPU costs come down, the number of tokens you eat goes up... the models are becoming cheaper and more efficient, but we're using more tokens at a faster rate and we're spending more." His conclusion is that software has to move to consumption or outcome-based pricing, and most SaaS companies aren't ready. The AI Files (Aug 23, 2026).

Not all tokens cost the same, by two orders of magnitude. On 20VC, the discussion pegged the gap between the premium frontier models and open-source alternatives at "double [digit] dollar cost per token" for the frontier versus "10, 11 cents per token" for open source, a difference big enough, the guest argued, to drive "massive adoption" of open models for everyday tasks while the expensive frontier models get reserved for the hard problems. The related warning: the companies renting out this compute aren't all equally healthy. Backing the inference provider Fireworks over rival Base 10, the guest's reasoning was pure operating discipline, calling it a "10 times better business, in my opinion, because they're more capital efficient" and actually willing to run at a profit rather than buy revenue at a loss. The Twenty Minute VC (20VC), "The AI Bubble Will Burst: Half the Neoclouds Will Die..." (Aug 22, 2026).

Your $100 subscription is being subsidized. A useful reminder for anyone budgeting off today's flat-rate plans: Keith Teare, who pays $100 a month for Claude and got throttled for the first time last week, estimated that to actually turn a profit on a heavy individual user "you'd probably have to pay $200 to $400 a month, maybe even more." Today's prices are a land-grab, not the steady state. Plan your unit economics for the bill you'll get after the subsidies end. Keen On America (Aug 23, 2026).