The company is becoming an eval harness

A startup this week launched what it describes as the first AI chief executive for fully autonomous companies: a system built from world models, recursive self-learning, a policy engine and a team of specialist agents, running a business end to end.

It already runs three businesses the company acquired. It says all three are performing, and points to a thread of numbers — one chart shows a cold-email campaign line climbing to $5,058. It is now putting **$2 million** behind a programme to acquire five more startups for the system to operate.

Strip away the launch language and something more interesting is happening than another agent demo.

The purchase is the benchmark

The standard way to evaluate an agent is a test set. The trouble is that test sets are written by the same people who build the agent, they leak into training data, and they measure the wrong thing: whether a model can answer, not whether it can operate.

Buying a company replaces all of that with a number that is produced by the world rather than by the vendor. Revenue, cost, churn and cash arrive monthly, in a bank account, whether or not anyone is watching. As one investor put it, acquiring distressed or stagnant software businesses and handing the keys to an agent harness is "the ultimate real-world eval" — the profit and loss statement becomes the objective function and the market becomes the feedback signal.

That is a different kind of claim from "our agent can use tools". It is a claim that the agent can hold a business steady while the market pushes on it.

Why stagnant software is the right substrate

The targets are deliberately unglamorous: small software businesses that already have customers, already have revenue, and have stopped growing. That matters for three reasons.

First, the product exists. The agent does not have to invent a market; it has to run one. Second, the cost base is mostly people — support, sales follow-up, maintenance — which is exactly what an agent stack can substitute. Third, the price is low, because a business with flat revenue and no story is worth little to a buyer who needs growth and a lot to a buyer who needs a test case.

What a P&L cannot tell you

The weakness is the same as the strength: an objective function is only as good as what it fails to measure.

An agent told to improve profit has an obvious path — cut support, raise prices, stop answering the difficult tickets — and each of those improves this month's number while damaging the asset. A human operator carries context the statement does not: reputation, goodwill, the customer who will leave in six months. Whether an agent learns to protect those, or simply harvests them, is the whole question, and a single quarter will not answer it.

There is also the base-rate problem. Three businesses reported as performing well is a sample of three, selected by the people reporting it, with no audited figures and no counterfactual: we do not know what those businesses would have done under their previous owners. The programme that doubles the sample to eight is the honest test — and even eight is small.

The structural shift

If the model works, the unit of acquisition changes. A company stops being only a cash flow and becomes a training environment with a scoreboard attached — and the cheapest ones, the stagnant businesses nobody wants, become the most valuable test cases in the market.

That is a strange place for the software industry to arrive at. For twenty years the exit for a small software business was another, larger software company. The new buyer may not want the product at all. It may want the operating record.

What to watch is not the launch thread. It is whether the acquired businesses publish audited results across several quarters, whether churn moves with profit, and whether the programme's next five acquisitions are bought at the same price. If profit rises and churn does not, the claim is real. If profit rises and the customers quietly leave, the eval was measuring the wrong number — which is the failure mode every benchmark has.

Source: company launch and investor commentary circulated on social media, 9–10 November 2026. Figures are the company's own claims and have not been independently verified.

Contact: [email protected]