The Horowitz Curve

The Horowitz Curve (n.) — The empirically observed decline in human task performance as a function of hours awake, plotted against the inconveniently flat line of an AI agent doing the same job.

We first noticed it doing code reviews.

At some point we were so tired the agent was a better reviewer. That got us thinking.

The model does not account for "I'll just check Slack real quick." Agent has no Slack. Task Performance Hours Since Waking 0 2 4 6 8 10 12 14 16 18 20 22 24 breakfast food coma afternoon coffee coffee kicks in coffee wearing off dinner asleep blood sugar dropping lunch the slow slide peak human The Horowitz Threshold Agent Human Fig. 1a: Task quality in humans vs. agents over 24 hours (n=everyone, p<you'd think so) Humans achieve 47% of agent output on a 24-hour basis. They describe this as "crushing it."

Now zoom out to a week.

Agents don't sleep. They don't take weekends. They don't need lunch. They do sometimes confidently produce nonsense, but so do humans.

Cumulative Output Day of Week Mon Tue Wed Thu Fri Sat Sun ~66% gap Agent Human Fig. 1b: Cumulative output over one week. The agent line is straight because it never stops. Humans achieve ~34% of agent output on a weekly basis. Weekends are not the flex you think they are.

Now zoom out to a year.

Agents don't take vacations. They don't get sick. They don't have slow Mondays. They will build the wrong thing faster than any human ever could.

Cumulative Output Month Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec ~69% gap Agent Human Fig. 1c: Cumulative output over one year. Vacations, holidays, and sick days are not in the agent's vocabulary. Humans achieve ~31% of agent output on a yearly basis. The agent doesn't even know it's a holiday.

And it applies to everything.

Code review was just the beginning. The Horowitz Curve shows up wherever humans perform tasks that agents can also do.

Code Review Driving Surgery Air Traffic Control Nuclear Launch Oversight Parenting Fig. 2: The Horowitz Curve has been observed across all fields.

It gets worse.

Humans aren't getting smarter. Agents are getting better every year. The window of human superiority isn't just narrow — it's shrinking.

Peak Task Performance Time 2022 2023 2024 2025 2026 2027 2028 Agents exceed peak human Human Agent Fig. 3: Frontier model performance vs. human performance over time. Once the agent line crosses, the Horowitz Curve from Fig. 1 has no window of human superiority. The entire 24 hours belongs to the agent.