THE NEED FOR SPEED

An attack is spreading through your network, and you want it stopped. But your agent sits waiting for your approval.

The hypegeist missed what a small fish can teach us about the cost of waiting.

On a reef in Curaçao, researchers filmed predators trying to catch small fish. The predators were bigger and faster, closing the gap until their prey was almost within reach. Then a last-moment turn would send the little fish out of their path. The predators’ nervous systems simply couldn’t respond in time. Their reactions took just 35 milliseconds, but even that was too slow for this fast food. Across the filmed attacks, the predators missed 97% of the time. And evolution seems to have little room to shorten that delay.

Microsoft’s latest Digital Defense Report describes AI shortening parts of cyberattacks from days to minutes. It argues that defenders will need to give agents more authority to respond.

In any case, keeping a human involved doesn’t necessarily mean they’re in control. At September’s congressional hearing on military AI, governance researcher Anna Mysyshyn asked asked what it means when one person supervises twelve autonomous systems. That person can remain responsible without the time to exercise that responsibility.

So how do you decide whether an agent should act or wait?

Getting it wrong has a cost, but so does waiting. So you need to weigh both.

Imagine a future with a million agents at work. You can’t approve every routine decision. And yet authority has to be earned through demonstrable improvements in the work.

The people who know the work must define what good decisions look like, set authority by decision type and establish enforceable limits.

There has to be a record of what people and agents decide, why, and what happens next. That evidence should determine whether authority widens or narrows, within the rules and as the work changes.

You remain responsible for the work. Where a decision needs your approval, make sure you have time to judge it.