Slow Agents Cost You More Than Time

The cost of an AI agent is not the minutes it takes to run, it is what you do with your attention while you wait.

Forty tabs

A podcaster I listen to described a stretch when he was busier than ever and still got almost nothing done. Agents that took one to three minutes per task left him with forty tabs open in his head, and the process he built to track them became a second job managing the first. This is not a willpower problem. Usability research summarized by Nielsen Norman Group says that after about ten seconds of waiting, people go do something else and then have to reorient. A short wait isn't short enough.

It is not just one person

Anthropic reported on October 16 that its agents sit idle around 80 percent of interactive session time waiting for a human, and that the share of time spent managing the work rose from 10 to 22 percent (Anthropic's research). I read that as the coordination tax showing up before the output does. (Anyone who has run a team of consultants knows the feeling. More threads rarely mean more finished work.)

The fix was speed

The fix is faster tools, so answers came back while you're still thinking. Fewer threads, more clarity, more depth. That's counterintuitive for those of us who talk about discipline first: fast tools let you be deliberate. It also separates two kinds of parallel work. Several agents can help with parts of one task. Parallel work across many unrelated tasks is the trap.

What to do

Time how long your team waits for an AI answer, and ask your leaders how many AI tasks they have open right now. Do you know the number for yourself?