Everyone wants to talk about AI agents Arun Devaraj

aiagentslimits

Everyone wants to talk about AI agents

The phrase is everywhere this year, so here is what it actually means, and where the excitement runs ahead of the machine.

by Arun Devaraj 3 minute read

Three different people asked me the same question last week, which is usually how I know a phrase has escaped the laboratory and got loose in ordinary conversation. This year the phrase is AI agents. It is in the headlines, in the adverts, in the mouth of someone's cousin who has just discovered it. So it seems worth doing what I always try to do here, which is to take the clever sounding thing and set it back down next to something ordinary until it stops looking like magic.

What the word is trying to say

For most of its short life, the kind of software everyone means by artificial intelligence did one thing at a time. You asked it something, it answered, and there it stopped, waiting for you like a very well read parrot. An agent is the name people have given to the same underlying machine when you let it take more than one step on its own. Instead of answering your question, it is meant to go and do the task: look something up, use the result to decide what to do next, reach for a tool, check the outcome, try again.

That is the whole idea, and it is a bigger idea than it sounds. A single answer cannot go wrong very far. A chain of steps can wander a long way from where you pointed it, because every step is built on the one before, and a small mistake early on is a large mistake by the end.

What they are good at

They are good at work that has a clear shape and a clear finish. Sorting a pile of documents into folders that already exist. Filling in a form from information sitting in an email. Taking a job you have described exactly and carrying it through the boring middle without needing you to hold its hand at each turn. When the path is well marked and someone can check the result at the end, this is real, and it saves real time.

Where the confidence runs ahead of the machine

The trouble starts when people hear go and do the task and quietly drop the last part: someone can check the result. An agent does not know when it is wrong. It has no small voice that says, this does not feel right, I had better stop. It will carry a wrong assumption all the way to a confident conclusion in exactly the tone it uses when it is correct. The machine is not being dishonest. It simply has no sense of its own edges.

So the more steps you let it take without looking, the more you are trusting not the cleverness of any single step but the whole chain holding together unwatched. That is a different and larger act of faith, and it is the part the excitement tends to skip.

The honest version

None of this is a reason to sneer at it. It is a reason to be precise. An agent is a fast, tireless, literal worker that never tells you when it has misunderstood you. Given a bounded job and a person who reads the output, that is a fine thing to have. Given an unbounded job and nobody watching, it is a way to make a mistake very efficiently.

If you take one thing from all this, take the plain shape of it. The machine has become able to keep going on its own. It has not become able to notice when it should stop. Those two are not the same skill, and almost everything worth understanding about this year's word lives in the gap between them.

More from the journal

All of the journal Every piece, by month

More by Arun Devaraj

Discussion

House rules

  • Sign how you like, but be one person. You may use the name on your account, a name of your own choosing, or none at all. What you may not do is wear somebody else's: no pen name of ours, no member's name but your own, and nothing that reads as the studio or its staff.
  • Argue with the point, not the person. Robust disagreement is welcome; contempt is not.
  • Nothing unlawful, hateful or harassing, nothing that identifies a private individual, and no advertising.
  • Post only what is yours to post. That includes other people's writing and private correspondence.
  • We may hide a remark, remove it, or close an account to the discussions, and we may do so without notice.

What a chosen name is and is not

A chosen name hides you from the room, not from us. Every remark stays linked to the account that wrote it, and the studio can see that link whenever it needs to. Please treat it as a way to speak freely, not as a way to say something you would not put your name to.

Where responsibility sits

Remarks in this thread are written by members and published without prior review. They are the views of the people who wrote them and not those of eQuill Studio. We do not verify their accuracy, and we accept no liability for them or for any decision taken in reliance on them.

If something here breaks these rules or infringes your rights, report it and we will look at it promptly.

This thread belongs to this piece alone. Members choose how they are signed, and can change it from the box below. When the studio or the writer joins in, the page says so with a badge, so you always know who is talking.

OPENING THE THREAD…