A unit of cognitive work output for the AI age.
One agentpower (AP) equals the total cognitive output of one knowledge worker, working full-time for one year, without AI assistance.
For 250 years, horsepower has been the universal unit for mechanical work output. James Watt invented it in 1782 not as a physics term, but as a translation layer: a way to help factory owners compare steam engines to the horses they already understood.
We need the same thing for cognitive work.
AI tokens are transforming the economics of knowledge work the way coal transformed the economics of physical work. But we're measuring the wrong thing. Token consumption tells you how much fuel was burned, not how much work was done. That's like measuring a factory by its coal bill.
Agentpower measures what matters: the total cognitive output of a human-AI system, benchmarked against what one unaided human produces.
Agentpower (AP)
Unit: 1 AP = the annual cognitive output of one full-time knowledge worker,
unaided by AI.
Scope: Measures the combined output of a human operator + AI tools,
not the AI alone.
Baseline: A knowledge worker without AI assistance operates at 1.0 AP.
AP Ratio = Total Cognitive Output ÷ Human Headcount
| AP Ratio | What It Means |
|---|---|
| 1.0 | No AI amplification. Pre-2023 baseline. |
| 2.0–3.0 | Early AI adoption. Routine tasks augmented. |
| 3.0–5.0 | Systematic AI integration across workflows. |
| 5.0–8.0 | Deep human-AI collaboration. Workflow redesign. |
| 8.0–10.0+ | Expert operator with full AI stack. Rare today. |
The old production function:
Output ≈ Labour
The new production function:
Output ≈ Labour × Fuel
Where Labour is human headcount and Fuel is AI token expenditure directed by human judgement.
Old model: 5 AP = 5 salaries → ~€310,000/year
New model: 5 AP = 1 salary + 4 × token budget → ~€62,064–€64,240/year
(€62,000 salary + €64–€2,240 in token fuel for 4 additional APs)
The additional agentpower (AP 2 through AP N) comes from tokens. The EUR/AP methodology shows the annual token cost for one full additional AP ranges from €16 to €560/year depending on model tier — so 4 additional APs cost €64–€2,240/year in fuel. Even at Claude Opus pricing, the new model is ~80% cheaper than five salaries. The human operator remains essential: they provide judgement, direction, context, and accountability. The AI provides scale.
See the EUR/AP pricing methodology for the full per-model breakdown and pricing derivation.
| Organisation | Headcount Change | Revenue Change | Implied AP Ratio |
|---|---|---|---|
| Klarna (2022→2026) | −49% | +104% | ~3.5 |
| AI-native SaaS companies (2025) | Lean teams | €0.9M–€3.7M rev/employee | 4.0–8.0 |
| Meta top engineer (2026, per CTO) | 1 person | ~10× normal output | ~10.0 |
| Pre-AI knowledge work baseline | — | ~€138K–€276K rev/employee | 1.0 |
Agent from Latin agere: to do, to act.
The word carries a double meaning that maps precisely onto this transition:
- Old meaning: A person who acts on behalf of another (lawyer, broker, consultant, administrator)
- New meaning: An AI system that acts autonomously to complete tasks
Agentpower measures what the two produce together.
The concept was introduced in the essay "Horsepower Then, Agentpower Now" published in Pale Blue Dot on Medium in April 2026.
The core argument: AI tokens are to knowledge work what coal was to physical work — a cheap, abundant fuel that transforms the economics of production from linear (output proportional to headcount) to multiplicative (output proportional to human judgement × token fuel).
Calculate your AP ratio. If it's 1.0, you're operating without amplification. Estimate what a token budget starting at €16–€112/year per knowledge worker (budget to mid-tier models) would do to your output. Even at premium frontier models, one additional full AP costs under €560/year in tokens — a fraction of a salary. That's your path to 2.0+. Use the AP Calculator to run the numbers.
Track AP ratios across sectors and regions as a leading indicator of economic transformation. Invest in training and institutional support — the constraint is no longer access to the engine, but the ability to operate it.
Your value is shifting from raw output to judgement, direction, and the ability to operate AI systems effectively. A professional at 5 AP is not five times busier. They're five times more capable.
We need better methods for measuring cognitive output across professions. The AP framework is intentionally simple — it's a starting point, not an endpoint. Contributions toward domain-specific AP benchmarks are welcome.
agentpower/
├── README.md ← You are here
├── LICENSE ← CC BY-SA 4.0
├── CITATION.cff ← How to cite this work
├── CONTRIBUTING.md ← How to contribute
├── references.bib ← Full BibTeX reference library (16 sources)
├── docs/
│ ├── essay.md ← Links to Medium publication (canonical)
│ ├── definition.md ← Formal definition and specification
│ └── faq.md ← Common questions and critiques
└── examples/
└── ap-calculator.md ← Simple worksheet for estimating AP ratio
@article{tom2026agentpower,
title = {Horsepower Then, Agentpower Now},
author = {Tom, Alphin},
journal = {Pale Blue Dot},
year = {2026},
month = {April},
url = {https://medium.com/alphin-pale-blue-dot/horsepower-then-agentpower-now-5bbeffabf3ef}
}This work is licensed under Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0).
You are free to share, adapt, and build upon this work, including for commercial purposes, as long as you give appropriate credit and distribute your contributions under the same license.
The term "agentpower" and the AP ratio framework are released into the public domain for universal use. No trademark is claimed or intended. The goal is adoption, not ownership.
See CONTRIBUTING.md. We welcome domain-specific AP benchmarks, translations, case studies, and critiques.
Agentpower is a project of Mycel UG, Berlin.
The engine doesn't have a moral compass. The operator does.