Tau Scaling Law

Computer programs that write texts or carry out tasks independently are often very reliable on short assignments. On long assignments, however, their success rate drops significantly. The Tau Scaling Law is a rule of thumb for how strongly it drops. The Greek letter tau stands in physics for a duration of time, and that is exactly what this is about: the length of a task. The rule states, roughly, that the probability of success halves once a task takes a certain amount of time longer. Instead of a fixed grade for a system, you get a statement of the form: tasks up to about an hour are usually managed, tasks over eight hours almost never.

Why task length is the more honest yardstick

For a long time, AI systems were evaluated using exam-style questions. They were given math problems, programming tasks, or multiple-choice tests, and the correct answers were counted. Each of these tests takes only seconds or minutes. They therefore say little about whether a system can survive an entire workday.

But that is exactly the economically interesting question. A company does not want to know whether a system can answer a question. It wants to know whether it can bring a project to completion. The task length a system can still manage is therefore a metric that can be directly compared with human labor.

Observations from recent years suggest that this manageable task length is growing quickly. Some analyses speak of a doubling roughly every seven months. Such figures should be treated with caution, since they depend heavily on the chosen test tasks. But they explain why investors and companies keep such a close eye on precisely this metric.

The chain of individual steps

At the core of the effect is simple probability calculation. A long task consists of many small steps performed one after another. Something can go wrong at every single step. For the overall task to succeed, practically all of the steps must succeed.

A numerical example illustrates this well. Suppose a system completes each individual step with 99 percent reliability. With 10 steps, it still gets through cleanly about 90 percent of the time. With 100 steps, that drops to only about 37 percent. With 500 steps, it is almost hopeless. The reliability per step has not deteriorated in the slightest.

Mathematically, this leads to a curve that runs flat at first and then drops steeply. The quantity tau indicates how quickly this collapse occurs. A common misconception is to understand the law as a natural constant. It is an observed regularity, not a proof. Moreover, tau depends on how tasks are cut and measured.

From chatbot to digital coworker

In news reports about AI, this idea usually appears without its name attached. When a provider announces that its new model now works independently on a programming task for several hours, this is exactly the metric being referred to. What is meant is not the computing time, but the time a human would need for the same work.

This is especially important for so-called agents. This is the term for AI systems that do not just answer, but independently operate tools, modify files, or search the web. Such systems work in long chains and are therefore precisely the application where the law bites hardest.

For practice, this leads to a clear strategy. Large assignments are broken down into short segments, and the result is checked after each segment. Every check resets the error chain back to zero. For this reason, AI agents are rarely left to run freely; instead, checkpoints are built in where a human or a test program verifies the work.