What Agents Can and Can't Do

Capability and reliability are not the same thing.

An agent can do things that would have seemed remarkable a few years ago. It can research a topic across many sources, write and run code, manage files, coordinate between systems. At their best, agents collapse tasks that used to take hours into minutes.

But agents inherit all the limits of the models underneath them, and then add new ones. Errors compound across steps in ways they don't in a single response. A mistake at step three, confidently built on, becomes a bigger mistake at step seven. The model has no instinct to stop and check.

Understanding agents clearly means holding both sides at once. What they can genuinely do, and where that usefulness ends. The distance between those two things is where most of the important questions live.