Model A and Model B: what comes after the 100:1 compression
The operational model I posted on April 23, 2005, expanded. Plus what the two emerging agentic-native startup archetypes look like; and what I'm building now.
On Wednesday, April 23, 2025, I posted this on LinkedIn:
The founder, VC and family office response: 159,689 impressions. 1,589 reactions. 373 comments. 117 reposts. The post made one structural claim: the next generation of large-employer-replacement companies will run on roughly ten senior operators, each owning a functional domain and managing their own stack of agents. Call it Model A. No junior headcount, no middle management, no SaaS wrappers. Capital profile: $10-12M up front, not $2M SAFE.
A few weeks back, Citrini Research published The 2028 Global Intelligence Crisis, painting a picture of the macro consequences of exactly the agent-mechanics I’d laid out: productivity without prosperity, SaaS margin compression, 10% unemployment by 2028, 38% drawdown in the S&P.
So let me pick that thread back up and tell you what I’m building.
A recap of the prediction I laid out in April 2025: the 100:1 compression
Every operationally complex company was built on management and specialist layers because coordinating work required human judgment at scale. Schedulers scheduled. Analysts analyzed. Developers wrote code. Auditors audited. Each layer of management and ERP systems was justified because the next layer up needed its output to do its job. This is essentially every Fortune 500 company that isn’t pure software, and most of the Fortune 1000 under them.
Agent-native companies collapse human layers. The senior orchestrator / architect stays, because taste and judgment matter more than ever. The thousands of executives, directors, managers, and specialists who used to translate senior judgment into daily execution are replaced by tuned agent stacks running at machine speed, machine scale, and structurally lower cost.
A year later, the claim has become the playbook. Enterprise software buyers are testing internal agent tooling in place of third-party SaaS. Incumbents across healthcare, retail, logistics, manufacturing, and financial services are running pilots where one senior operator plus an agent stack does the work of an entire team. The April 2025 post described what will replace the Fortune 500 operating model over the next decade, and what a new generation of ten-person-led operating companies will be built on from day one.
Why corporate America will adapt
The second lever implicit in my April 2025 post was competitive pressure on incumbents. If a new generation of ten-person-led agent-native companies can do what a 1000 person legacy operations org does, that incumbent is going to restructure toward the same shape as the attacker. The incumbent CHRO doesn’t need to believe the thesis; they need one peer CEO to hit a margin target the legacy structure can’t reach, and it’s restructure or fail. Greenfield companies will move faster. The incumbents are bigger ships to turn, but it’s going to be turn or sink. Citrini’s macro picture, productivity surging while employment contracts, is what happens when both run at the same time.
What I’m building next
My April 2025 post was about Model A, the ops-heavy company. There is a second model that matters just as much, and that I’ll call Model B: the pure software company with 1-2 founders running a disciplined agent harness.
I’m working on a new company in each model, both launching this summer.
The Model B company: Project Darkhouse
Model B is the pure software version of the same order-of-magnitude compression of human coordination layers. One founder, sometimes two, running a disciplined agent harness against a structured vault. The founder holds the judgment layer, and agents execute in scoped lanes with formal handoff gates. The architecture is the discipline. A founder running this pattern in 2026 ships what a whole well-funded team shipped in 2020. Implications for pandemic-era SaaS vintages are they don’t get to keep their cost structure when a bootstrapped AI-native competitor arrives in the same category, ships the same workflow tool in a fraction of the time.
There’s a whole long tail of categories that were never economical at venture scale, that are now within reach. What I like to call the new Mittelstand moment. Businesses with a few thousand to a few hundred thousand serious users, 7 to 9 digit TAMs, real demand, real willingness to pay, real churn protection through hillclimbing domain expertise. A generation of real businesses sat in plain sight, previously uneconomical. That’s an enormous surface area the last decade of venture couldn’t till, and it sits almost entirely in consumer and small business for the obvious reason that enterprise belongs to Model A startups, or to incumbents building their own internal agent stacks.
Project Darkhouse is a vehicle for launching several of these niches into prosumer opportunities. Millions of adults in America compete in sports outside of work. The same knowledge gap exists in every one: context is trapped in manufacturer spec sheets, in coaches and pro-shop techs, in word of mouth, and in forum threads. What comes next is founders putting together world-class domain context, and then setting agents loose on top of it. Putting world-class tooling in the hands of prosumer practitioners and coaches. Building in Minneapolis and shipping this summer. First niche is one with 3.2 million US competitive enthusiasts, category-wide knowledge gap, zero good structured reference product. More categories to follow.
My bet on Model A: a large-corporate attacker
I’m also building an attacker in a 50+ year old industry. High operational complexity, hierarchy-heavy legacy, exactly the kind of incumbent structure agent-native companies are positioned to replace. That’s a separate post when we’re closer to launch.
What this means if you’re a founder
For Model A, the opportunities are the large operating incumbents whose structure is vulnerable to an agent-native attacker: the ERP-era operating models, the thousands of executives, directors, managers, and specialists, the cost structures built on an assumption of human coordination that no longer holds. The founder retired years ago, and the AI response will be slow.
For Model B, the opportunities are the categories software couldn’t effectively reach previously: professionals and prosumers with real willingness to pay in markets too small for a 15-person team to close the math, and domain knowledge trapped in manufacturer sheets, coaches, techs, forums, and veteran memory. You find those by hunting for practitioner communities with real depth and no structured infrastructure to support them. Wherever the state of the art lives in people’s heads, there’s a Model B company to build.
Evaluating whether what you’ve found can become a monopoly matters as much as finding it. World-class software and cap table are becoming less central as moats. Model performance and agent stacks will commoditize. What will compound is proprietary organizational context, workflow embedding, and switching costs that accumulate through use. Ask yourself three questions about any opportunity (after Thiel and Helmer): Does the business accumulate context through use that no competitor can replicate from a cold start? Does it sit inside a workflow the user would rebuild painfully if they left? Does every new user or transaction make the next one more valuable? One or two, you have a good business but not a category-defining one. Zero, you have a feature. Three yes answers, you have a monopoly-shaped opportunity.
What this changes for capital
Both models break traditional early-stage VC.
Model B because venture’s underwriting stack wasn’t built for multi-bet solo founders. The round the economics would support is below venture’s minimum viable check. The founder typically won’t raise until revenue already works, which means no traction gap for venture to close. Underwriting has to shift from team size and burn rate to agent-stack quality, accumulated organizational context, and revenue durability. Most firms aren’t set up to evaluate those dimensions.
Model A breaks it because the team is too senior, the check is too big, the plan is too real for early-stage venture to underwrite. The $10-12M up front goes almost entirely to compute and senior operator salaries, and the capital is as likely to come from operator-investors who ran the pre-LLM version of the business as it is from private equity or growth-stage venture.
Net-net, venture as currently structured can’t really serve either model well. Model A wants capital but not at venture’s check sizes or governance terms. Growth equity and operator-LP funds are closest to the shape these companies need; traditional early-stage VC is not. Model B wants capital below venture’s minimum viable scale, or doesn’t want it at all. Venture will need to restructure to serve these companies, or it will capitalize the shrinking universe of companies that still fit the old pattern.
The window
The April 2025 post was Model A. Model B is the other half of the picture, and the two archetypes together are how operating companies and software companies will be built for the rest of the decade. For founders, the window is open right now. The mechanics are live, categories are visible, and the consensus is still forming. Grab a problem that speaks to you. Or grab two or three, because Model B economics mean you don’t have to pick just one. If Darkhouse takes 40% of my time, I can run a second bet with the other 40% and still have 20% for compounding infrastructure. The founder who would have built one company now builds three. Someone else is already or is about to plant their flag on them if you don’t.