The word "agent" is doing a lot of work in AI marketing right now, usually while standing on a table wearing a superhero cape. Product launches promise digital employees, autonomous workers, companies of one running on fleets of them. Strip the cape off and the concept is simple enough to explain to a colleague in one minute — and once you have it, you'll be better at deciding when to use an agent than most of the people selling you one.
The gap between the concept and the marketing matters, because agents are genuinely useful and genuinely oversold at the same time. Knowing which half a task falls into is becoming a practical workplace skill. It starts with one distinction.
Chatbots answer, agents act
A chatbot does one thing: it takes your text and returns text. Ask it a question, get an answer. That's the whole interaction.
An agent is the same model plus tools plus a loop. The tools might be a browser, a file system, a calculator, an API to your calendar — things that act on the world instead of describing it. The loop is the plan–act–check cycle: the agent figures out what to do, does it, looks at what happened, adjusts, and repeats until the task is done or it gives up.
The difference in practice: you can ask a chatbot to explain how to find the cheapest flight, and ask an agent to go find it. One describes; the other operates.
The useful analogy: an intern with a computer, no judgment
Picture a capable intern who types fast, reads fast, and never gets bored — and who has no judgment about any of it. That's roughly what you're delegating to when you hand work to an agent. The comparison isn't an insult; most valuable work in any office was done by people exactly like this under supervision. The supervision is the part you still supply.
The analogy predicts the strengths. Give the intern bounded, checkable, repetitive work and they shine: drafting and filing documents, pulling data from five places into one spreadsheet, running through a checklist of fifty steps. Tasks where the goal is clear, the steps are enumerable, and you can verify the result without redoing the work.
It also predicts the failures, which is the part the demos skip.
Where agents go wrong
Three failure modes matter, and they compound.
Errors compound across steps. An agent that's 95% reliable per step is roughly 60% reliable over ten steps and under 36% over twenty. Long tasks aren't long versions of short tasks; they're multiplication problems, and the product shrinks fast.
The letter, not the intent. An agent optimizes exactly what you instructed, not what you meant. Ask it to "reduce costs" and it will find a way to reduce costs — possibly the one that deletes something you needed. This isn't stupidity; it's obedience, taken literally.
It can't be accountable. When the outcome matters — money, legal exposure, someone's data — a trail of agent actions doesn't substitute for a person who answers for the result.
Rules for actually using one
A few rules keep the intern useful:
- Keep tasks small. Ten steps with verification beats a hundred-step autonomous run.
- Build in real checkpoints — points where the agent stops and you look at what it did before it continues.
- Put human review at every step that carries consequences: anything involving payments, deletions, external communication, or decisions about people.
- Prefer tasks where checking the output is faster than doing the work. If verification is the hard part, you've delegated the wrong half.
None of this makes agents a fad. Bounded, checkable, repetitive work is a large share of most jobs, and agents are getting genuinely better at it. But agents are a delegation tool, and delegation was never free: you manage what you delegate. The productivity comes from choosing what to hand over — not from handing over everything and hoping.
No comments yet