The dominant metaphor for agents is the new hire. Give it access, give it context, let it learn the ropes, promote it to bigger tasks as trust builds. It is an appealing picture and it mispredicts almost every failure I have seen.
Here is the thing the metaphor gets wrong, and it is not a small detail.
A junior engineer who does not understand a task becomes visibly stuck. They re-read the ticket. They ask a colleague. They stall, they hesitate, eventually they escalate. That stall is not a defect in the junior engineer — it is a safety mechanism, and every process we have built around software teams quietly depends on it. Standups exist to surface stalls. Code review catches the work of someone who proceeded without understanding. The whole apparatus assumes that confusion produces a signal.
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, and the first place anyone notices is wherever the work lands.
What follows
Verification cannot be sampled. With people you review a fraction and trust the rest, and that is sound statistics: the failure rate is low, failures are correlated with signals you can see, and people tell you when they are unsure. None of those three hold here.
Scope matters more than capability. The question I have learned to ask is not whether the agent can do a thing. It is: if it does this wrong, what does that cost, and when would I find out? Cheap and immediately is where these tools are excellent. Expensive and eventually is where I will not put them, no matter how good the demo was.
Give it the ability to be checked. In a hackathon we tried a local model as an LLM-as-a-judge for validating a handful of STIG rules — ten or fifteen of them, deliberately a small and checkable set. The reason that experiment was worth anything is that every judgement it made could be held against a real system and confirmed or thrown out. An agent that produces an answer plus the evidence for the answer is worth several that produce answers alone.
Where the metaphor does hold
Context is everything. Nearly every disappointing result I have had was under-briefed rather than under-capable — the same failure as handing a new person a ticket with no background and being unimpressed with what comes back.
The difference is that the new person tells you. That difference is the entire design problem, and the reason I keep a human gate on anything that leaves my hands.