More AI agents per worker. More work per agent.

Every action is one reading, paid per action, for the deployer, the insurer and the regulator alike, so whoever pays for the reading is earning far more from the action.

The only thing you have to believe is what the market already believes: more agents per worker, and more work per agent. Multiply the two and you get this curve.

What the curve is: the yearly value of reading every agent action at our list price, from the Agent-Year Meter’s model. The licence is $20 per agent-year, covering 10,000 attested actions or 365 days, whichever comes first.

Sourced: hyperscaler capex of $775–800 billion in 2026 (AL Capital Advisory · Futurum). The grey line’s growth rates are the lowest on record for each driver: 3.1% a month for the share of workers with agents and 3.1% for agents per worker (the agentic-AI market’s 43.6% a year), 6.8% for actions per agent (the industry’s own 24× token projection to 2030). The gold line is the rate the capex is priced on.

Assumed: where it starts. 245 million workers (7% of 3.5 billion) with 1.5 agents each, 100 actions per agent a day.

Stress test: start ten times smaller and the grey line still passes $1 trillion in Sep 2032. The date moves; the shape does not.

Every assumption is a slider in the full model, with its source, in the data room.

The bad case is not rare. On our own agents, 1 task in 5 ran past twice the typical task’s tokens and 1 in 18 past four times; the worst used 310.4M. That is the tail an underwriter has to price, and it is bad news for any story of control.

With the job declared before the task: 46% fewer tokens for the same task mix (permutation p = 0.005), and the worst task 15.4M against 310.4M (p = 0.02). n = 44 declared, 598 not.

Our own agents, and not a randomised trial: the two groups are clearly different, and that does not prove the declaration caused it. Live: thetadriven.com/tail

Since 1 January 2026, standard liability forms let insurers exclude generative AI (ISO CG 40 47, CG 40 48, CG 35 08; Fenwick), and more than 60 insurance groups have filed to adopt them (The Insurer, 23 July 2026; Insurance Journal, 17 August 2026). If it’s debatable, it’s not insurable.

We sell the Richter scale: how far each agent goes past the job it was given, read from what it actually changed. Detected · placed · priced · dispatched. You hold the wheel.

Ask an AI claims agent whether it did its job and it will say yes. That answer is made of the same words as the work, so it cannot settle anything.

Think of a plumber who, to finish the job, starts doing electrical work. You may be fine with that or not, but you want to know it happened. A reading of what actually changed shows it; asking the plumber does not.

A Richter scale does not stop an earthquake, and a flight recorder does not stop a crash. They make the operator responsible, and responsibility is what gets insured. There were no highways before driver’s licences and insurance.

Why not just set rules, like “approve nothing over $X”? Then it is an if-statement, not an agent. The decisions worth delegating are judgement calls, and a rule either lobotomises the agent or misses them.

Why not have another AI check it? A better judge does not help: whether software behaved is undecidable (Rice, 1953), and an account written by the process cannot contain what the process left out. The only move left is a record the actor did not write.

Won’t the labs or the brokers build it in-house? They can build the arithmetic. They cannot build independence: a measurer paid by the thing it measures cannot certify it. Hartford Steam Boiler did not make boilers; it inspected them, and that is why the inspection counted. A broker who says it will build this in-house is telling you it needs it; the measurement is open so brokers and carriers can build their own risk products on it, and the licence is what makes a reading count.

Does it stop an agent doing harm? No, and it does not claim to. It detects, places and prices; you decide how strict to be and wire the halt.

Is the maths solid? It is a closed algebraic identity: the proof closes, so there is nothing to negotiate in the arithmetic. What you can argue about is your specification, which is the point: it is yours.

Hardware or software? It ships today as open-source software that runs on your own machine and gives the same answer on every rerun. Binding it to the chip is what the patent covers, and that is pending.

What is not proven yet? No carrier has priced on it yet, the patent is not granted, and the tail above is our own agents, not a randomised trial. We say so here so you do not have to find it.

A tool competes on features. A unit does not: once a premium is quoted per agent-year, every deployer, broker and carrier quotes in it. That is why the measurement is free (MIT): it has to be everywhere before it can be the unit.

Patent pending, not granted: US 19/637,714, filed 2 April 2026 under Track One, 36 claims.

Agents × $20 per agent-year, the list price. Arithmetic, not a forecast; the model behind the curve above is in the data room.

Measuring is free and open-source (MIT), and stays that way, so it spreads with no sales call. What is paid is the third-party attested backup an insurer will accept.

For your engineer, the check takes a minute: npx -y thetacog-mcp@latest attest-demo, run twice. Same bytes, same hash.

If you invest, the data room is by request.