Skip to content
MycelLet's Talk
FRAMEWORK · PUBLIC DOMAIN · APRIL 2026

Agentpower

The unit of cognitive work output for the AI age.

A brass industrial gear with luminous teal threads radiating outward — the Agentpower hero image

1 AP = the total cognitive output of one knowledge worker, working full-time for one year, without AI assistance.

In 1782, James Watt needed to sell steam engines to brewery owners. The brewers didn't understand pressure differentials or thermal efficiency. They understood horses. So Watt invented a word: horsepower. One engine, he told them, produces the equivalent of twenty horses.

The brewers bought the engines.

Two hundred and forty-four years later, AI tokens are transforming knowledge work the way coal transformed physical work. But we're still 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.


The AP Ratio

1.0
Baseline
No AI amplificationAll knowledge work before 2023
2.0
Early
Selective AI use. Routine tasks augmented.Most organisations, 2026
3.5
Integrated
Systematic integration. Roles redesigned.Klarna 2026 — −49% staff, +104% revenue
5.0
Advanced
Deep human-AI collaboration.AI-native SaaS · €1M+ rev/employee
10.0
Expert
Full AI stack. Token budget ≈ salary.Meta top engineer · per CTO Bosworth, 2026
mycel-ai.de/agentpower · CC BY-SA 4.0

The AP Ratio measures the average amplification factor across a team, department, or organisation:

AP Ratio = Total Cognitive Output ÷ Human Headcount

An AP ratio of 1.0 is the world before 2023. Klarna runs at roughly 3.5. Meta's CTO recently said his most productive engineer produces roughly ten times normal output while spending the equivalent of his salary on tokens — a personal AP ratio of 10.


The Production Function Shift

Then
OutputLabour
Linear. Revenue tracks headcount.
1 hire → ~2× revenue. Proportional growth.
Now
OutputLabour × Fuel
Multiplicative. Token fuel amplifies each worker.
1 hire + tokens → 5–10× output.
1× baseline
5–10× amplified
mycel-ai.de/agentpower · CC BY-SA 4.0

More heads, more output — that was the old formula. Proportionally, predictably, linearly. It shaped how organisations were structured, how companies were valued, how entire economies were planned.

That formula is breaking.

Where Labour is human headcount and Fuel is AI token expenditure directed by human judgement. The relationship is multiplicative, not additive. Tokens without human direction produce noise. Humans without tokens produce at baseline. The product of the two is where the amplification happens.


The Economics

In the old world, one agentpower costs what one employee costs. The median EU knowledge worker costs roughly €62,000 per year all-in (salary plus employer contributions).

In the new world, additional agentpower comes from tokens — at a fraction of that cost.

Old Model
€310K
5 salaries / year (EU median)
Salary ×1
Salary ×2
Salary ×3
Salary ×4
Salary ×5
New Model
~€64K
€62K salary + €64–€2,240 token fuel / year
Salary ×1 (the operator)
Tokens
~80% structural cost reduction
mycel-ai.de/agentpower · CC BY-SA 4.0

The first AP still costs a salary — you need the human operator. Each additional AP above that costs only token fuel. Even at frontier model pricing, the new model is ~80% cheaper than five salaries.

This is the structural economic shift. Not an incremental improvement.

Annual cost of one additional AP · SCW 500K tokens/day × 224 EU working days · €1 = $1.09
DeepSeek V3.2DeepSeek
16
DeepSeek V3.2 ReasonerDeepSeek
23
GPT-5.4 nanoOpenAI
26
MiniMax M2.7MiniMax
29
Gemini 3.1 Flash LiteGoogle
32
Gemini 3 FlashGoogle
64
GPT-5.4 miniOpenAI
96
Claude Haiku 4.5Anthropic
112
GLM 5.1Zhipu AI
122
Gemini 3.1 ProGoogle
255
GPT-5.4OpenAI
319
Claude Sonnet 4.6Anthropic
336
Claude Opus 4.6Anthropic
560
1 human hireEU median, ISCO-2
€62K+
SCW: 500K tokens/day (150K input + 200K cached + 75K output + 75K reasoning)
× 224 EU working days/year · Human ref: €48K median gross + ~30% employer costs
mycel-ai.de/agentpower · CC BY-SA 4.0

Split image: Victorian steam horsepower gauge on the left, AI data streams from a modern screen on the right
Horsepower then. Agentpower now.
Then
Horsepower
1782
Now
Agentpower
2026
Named after
The thing being replaced (horse)
The role being amplified (agent)
Measures
Mechanical work output
Cognitive work output
Fuel
Coal
AI tokens
Baseline unit
1 horse, 1 year
1 knowledge worker, 1 year
Multiplier
~20× per steam engine
3–10× per human-AI system
Cost advantage
Steam at 25–30% cost of horses by 1850
Marginal AP at ~3–5% cost of additional hire
Key insight
Measure output, not coal burned
Measure output, not tokens consumed
Fuel access
Geographically fixed. Coal requires mines and railways.
Globally distributed. Tokens require only internet access.
mycel-ai.de/agentpower · CC BY-SA 4.0

Real-World Reference Points

| Organisation | Change | Implied AP Ratio | |---|---|---| | Klarna (2022→2026) | −49% headcount, +104% revenue | ~3.5 | | AI-native SaaS (2025) | €0.9M–€3.7M revenue per employee | 4.0–8.0 | | Meta top engineer (2026) | ~10× normal output (per CTO) | ~10.0 | | Pre-AI knowledge work baseline | — | 1.0 |


Calculate Your AP Ratio

mycel-ai.de/agentpower · CC BY-SA 4.0
Your AP Ratio
1.5
Early adoption
Some tasks augmented. Significant room to grow.
123.5510
Cost of your next AP
Via hiring
62,000
Via tokens
300
Cost ratio
207×

At 5× or more the economics favour AI augmentation structurally, not marginally.


Why "Agent"?

The word agent holds two meanings that map precisely onto this transition:

  • Old meaning: A person who acts on behalf of another — the lawyer, the broker, the consultant, the administrator
  • New meaning: An AI system that acts autonomously to complete tasks

Agentpower measures what the two produce together. The human is the operator. The AI is the engine. Without the human's judgement, direction, and context, the engine produces noise. Without the engine, the human is limited to 1 AP.

The power is in the pairing.

The word comes from Latin agere: to act. It translates naturally across languages. No trademark is claimed. The goal is universal adoption.

Victorian brass pressure gauge with luminous teal dots on the scale — visual metaphor for the Agentpower ratio
The gauge metaphor: an old instrument, a new scale.

Scope and Limitations

Agentpower applies to cognitive work — any professional activity where the primary output is produced by thinking, analysing, writing, communicating, calculating, diagnosing, designing, advising, or deciding.

What it does not measure:

  • ·Token consumption (input metric, not output metric)
  • ·Revenue per employee (conflates pricing, market conditions, and productivity)
  • ·Hours worked (measures time, not output)

Honest limitations:

Cognitive output is hard to measure. Unlike horsepower, which maps to a precise physical quantity, cognitive output varies by profession and context. The AP framework is directional, not exact.

Not all tasks amplify equally. Routine, structured tasks amplify more readily than novel, creative, or interpersonal tasks. Quality matters — 5 AP of mediocre output is not 5 AP.

Token costs are falling. The economics described here reflect 2026 pricing. As costs continue to decline, the marginal cost of additional AP will decrease further.


Frequently Asked Questions

Why do we need a new unit? Can't we just use 'revenue per employee'?+
Revenue per employee conflates too many variables: pricing power, market conditions, product mix, geographic cost differences. It tells you something about a business but nothing specific about AI-augmented productivity. Agentpower isolates the question: how much cognitive work does each human on your team produce, relative to what they'd produce unaided?
Isn't this just 'productivity' with a new name?+
Productivity is a broad economic concept measured in many ways. Agentpower is specific: it benchmarks AI-augmented cognitive output against a defined human baseline. The specificity is the point. 'Productivity' is too vague to drive decisions. 'Our AP ratio is 1.2 and our competitor's is 3.5' drives decisions.
How do you actually measure cognitive output?+
This is the hardest question, and the honest answer is: imperfectly, for now. Different professions will need domain-specific proxies. For lawyers: contracts drafted, cases managed, filings completed. For engineers: calculations run, designs reviewed, specifications written. The framework is deliberately simple at this stage to encourage adoption. Better measurement methodologies will follow.
Can you game the AP ratio?+
Yes, the same way you can game any metric. That's why agentpower explicitly measures output, not input. Tokenmaxxing — maximising token consumption as a proxy for productivity — is the wrong metric for the same reason that measuring coal consumption was the wrong metric. Agentpower resists this class of gaming by anchoring to output, not fuel.
Does a high AP ratio mean fewer jobs?+
Not necessarily. A high AP ratio means more output per person. Historically, productivity gains have expanded total economic output rather than shrinking total employment, though they do change which jobs exist. The steam engine eliminated some roles and created others that were unimaginable before it.
Is this only relevant for large companies?+
The opposite. Large companies have always been able to scale by hiring. The AP ratio matters most for smaller organisations, because it means a team of 10 can now produce the output of a team of 40. That's a structural advantage for lean, well-run organisations that couldn't previously compete on scale.
Is 'agentpower' trademarked?+
No. The term is explicitly released into the public domain. The goal is universal adoption, not ownership. Following the convention of horsepower — one word, abbreviated as AP, lowercase in running text.

Origin and License

The concept was introduced in the essay Horsepower Then, Agentpower Now published in Pale Blue Dot on Medium in April 2026. The full definition, methodology, and EUR/AP pricing derivation are available on GitHub under CC BY-SA 4.0.

The term agentpower and the AP ratio framework are released into the public domain for universal use. No trademark is claimed. The goal is adoption, not ownership.

The engine doesn't have a moral compass. The operator does.

Published as an open resource — read, share, use freely.