What AI Agents Still Cannot Do (Notes From a Year of Trying to Delegate)

I have spent a year trying to actually delegate work to AI agents: real tasks, real deadlines, real consequences. The demos are not fake; the tasks in them do get done. What the demos hide is the shape of the boundary, which has turned out to be remarkably consistent across every agent I have tried.

The first side of the boundary: agents fail at goals whose completion is ambiguous. 'Clean up this codebase' produces confidently wrong reorganization, because there was no spec and the agent cannot ask the office questions that would surface one. Delegation works exactly to the degree that done is definable in advance. Most knowledge work is not.

The second side: agents fail at contexts that resist enumeration. Booking a flight is enumerable (sites, forms, confirmation). Renegotiating a contract with a person who is having a bad day is not. The failure is not capability; it is that real contexts contain information that never makes it into any prompt, and the agent does not know what it does not know.

The third side is the quiet one: agents fail at consequence. They will act on a wrong premise with the same confidence as a right one, because nothing in their loop feels stakes. Every real delegation system I built eventually grew a human checkpoint at exactly the moment of irreversible action, not because the agent erred often, but because when it erred there was no signal until the damage.

None of this is a dismissal; the year also produced real wins in the enumerable-and-definable zone, which is bigger than skeptics think. But the honest map of the boundary matters more than the demos, because the failed delegations were not capacity problems. They were specification problems, and no amount of model improvement writes your specification for you.