The dominant metaphor for agents is a new hire: give it access, give it context, it learns the ropes. It is an appealing picture and it mispredicts almost every failure I have seen.
A junior engineer who does not understand a task becomes visibly stuck. They ask, they stall, they escalate. That stall is a safety mechanism and the whole organisation is built around it.
An agent that does not understand a task produces confident, plausible, complete-looking output. There is no stall. The signal your process depends on simply does not arrive.
What follows from that
Verification cannot be sampled. With people you review a fraction and trust the rest, because the failure rate is low and failures announce themselves. Neither holds here.
Scope matters more than capability. The right question is not "can it do this" but "if it does this wrong, what does that cost and when would I notice". Tasks where the answer is cheap-and-immediately are where agents shine.
Give it the ability to be checked. An agent that produces an answer plus the evidence for the answer is worth several that produce only answers.
Where the metaphor does hold
Context is everything. Most disappointing results I have seen were under-briefed rather than under-capable — the same failure mode as handing a new person a ticket with no background. The difference is that the new person tells you.