hello

The power law of practice is absent from agents

San Francisco, California • Aug 6, 2026

There's a law that governs almost everything humans learn to do. It's called the power law of practice: the time it takes to perform a task falls predictably as you repeat it [1]. The hundredth time you tie your shoes, sign your name, or run a familiar report, you get faster, more efficient, and more predictable.

This law manifests in all occupations: typists, surgeons, factory lines, chess players, you name it. It’s one of the most robust regularities in all of human performance [2]. But it is almost entirely absent from the way the industry builds AI agents today.

That absence creates problems. AI agents, as the industry currently builds them, lack efficiency.

To be clear, this isn’t about AI’s capability; frontier labs have been breaking new grounds virtually every month. The problem is agent cost – measured in dollars, time, tokens, compute, and human oversight – and the fact that none of those costs gravitate toward the most efficient result. Just as how things in nature tend towards efficiency, computer autonomy should, too.

Why the curve is flat

If agents should get cheaper with practice, why isn’t the industry doing that right now? The reasons are threefold.

First, economics rewards the opposite. The prevailing economic model of the token business is capital-driven and built around APIs. That means the more tokens users burn, the more the providers earn. The incentive doesn’t push for efficiency but shapes the product into longer chains, bigger context, 'think harder' by default. A power law of practice would bend the cost line down over time, but the business model needs it to stay flat or climb.

Second, the interface is built for conversation. Agents today are built as chat interfaces, but chat is fundamentally at odds with the nature of work. The purpose of work is to turn effort into repeatable systems. Repeatability builds muscle memory, muscle memory reduces cognitive load, and reduced load is what produces efficiency. Chat does none of this. Work and chat are structurally different, and the interface for a work agent should not be chat.

Lastly, today's agents don’t practice because they never learn on the task. Agents of today are predominantly LLM-based. While an LLM model is great at solving novel problems, it’s wasteful on routine tasks. It doesn’t learn on the task but re-inferences from scratch every time. The LLM keeps doing the work without ever forming muscle memory, so it never gets more efficient at a task it has already done a hundred times. This turns into billable amnesia.

What practice looks like in a machine

So how do we build agents that actually follow the power law of practice? The good news is we already have a working example of extreme efficiency sitting inside our brain.

The human brain runs on about 20 watts [3]. That is the power of a dim light bulb, and with it, a person can hold a conversation, recognize a face across a noisy room, plan a week of meals, and second-guess a decision from 10 years ago. No machine we have built comes close to that efficiency, namely the ratio of output to energy consumption.

It’s worth sitting with that before talking about machine intelligence. The brain did not become efficient by being powerful. It gets there by being cheap, and it gets cheap through practice.

An AI agent should work the same way. It shouldn’t pay twice for the same task, and it shouldn’t starve the model to save on tokens. Like muscle, learned and repeated actions can be handled by a small model and stored in small memory units. Not everything needs to route through a central LLM – the "brain" – every time. Intelligence doesn’t require brute force; small, structured, symbolic units can execute repeated work in code efficiently. That is the power law of practice: the first run is an expensive discovery, and every run after gets cheaper.

This is also the neurosymbolic approach: neural LLMs where novelty is needed for discovery, and symbolic, muscle-memory-like code for repeatable work. The neural half handles first encounters; the symbolic half is where practice happens. Together, they give agents efficiency overtime.

Efficiency is the point, not token cost

Albeit the industry’s focus on token-maxxing or -minimizing, neither is the point. The point in question is how to improve agents with practice, like humans going to the gym doing their routine workout. Efficiency becomes your moat rather than AI labs' revenue line. They want the token counter to go up. An agent system that practices brings it down.

References

[1] Newell, A., & Rosenbloom, P. S. (1981). Mechanisms of skill acquisition and the law of practice. In J. R. Anderson (Ed.), Cognitive Skills and Their Acquisition (pp. 1–55). Erlbaum. The foundational formulation of the power law of practice. See also Snoddy, G. S. (1926), "Learning and stability," Journal of Applied Psychology, 10(1), 1–36, an early empirical observation of the effect.

[2] The power-law form is widely replicated across domains (typing, mental arithmetic, motor tasks, cigar rolling, etc.). Note the ongoing debate over its exact shape: Heathcote, A., Brown, S., & Mewhort, D. J. K. (2000), "The power law repealed: The case for an exponential law of practice," Psychonomic Bulletin & Review, 7(2), 185–207, argue that individual learning curves are better fit by an exponential than a power law — but the core finding, that performance improves systematically with repetition, is not in dispute.

[3] Estimates of the adult human brain's power consumption cluster around 20 watts (~20% of the body's resting metabolic rate). See Raichle, M. E., & Gusnard, D. A. (2002), "Appraising the brain's energy budget," PNAS, 99(16), 10237–10239.

Building autonomous computers doesn't mean replacing humans. It means cooperation.

Free your hands from the computer. Download Simular today for free.

Try Sai