What is an AI agent actually doing?

The setup

An agent isn't a smarter app. It's a different shape.

Regular software is a recipe. A developer wrote down every step in advance — "when a form is submitted, save these fields, send this email, show this page." It does exactly that, every time, in that order. Change the situation and it can't adapt; it only knows the steps it was given. That's most of the software you've ever used, and for most jobs it's exactly what you want.

An agent is built the opposite way. You don't hand it the steps — you hand it a goal, a set of tools it's allowed to use, and you let the AI decide the steps as it goes. Think of the difference between a paint-by-numbers kit and a finish carpenter you turn loose in a room. The kit can only produce the one picture it was printed for. The carpenter you give a goal — "trim out this room" — and he sizes it up, picks the tools, works in the order that makes sense, and adjusts when a wall isn't square. The agent is the carpenter, not the kit.

That's the load-bearing line, so it's worth saying flat out: in regular software the developer controls the path; in an agent, the AI controls the path. Everything else about agents falls out of that one shift.

The mechanism

The loop it's actually running.

Strip away the buzzwords and an agent is doing the same small cycle over and over until the job is done. Four steps:

That's it. Think, act, look, repeat — until the goal is reached. A human assistant clearing a task does the same thing without naming it: tries something, sees how it went, adjusts, tries the next thing. The "intelligence" people imagine isn't a single magic answer — it's the model getting to course-correct across many small steps, because each loop it can see how the last one turned out.

Why it's useful

What the loop buys you.

The reason this shape matters is that real work is messy and rarely goes in a straight line. A rigid script breaks the moment reality doesn't match what the developer pictured. The agent's loop is built to bend instead of break, and that gives you three things plain automation can't:

• It handles the "it depends" jobs. Tasks where the right next step changes based on what you find — chase the record, and if it's missing, look somewhere else — are exactly where a fixed script falls apart and a loop shines.
• It recovers instead of stopping. When a step fails, the agent sees the error and can try another way — a different tool, a corrected input — rather than dying on the spot the way a script does.
• It strings many steps together. One request — "pull this week's new filings, match them to our list, draft the follow-ups" — can become a dozen tool calls the agent sequences itself, instead of you wiring every one by hand.

None of that requires the AI to be a genius. It requires the AI to be allowed to keep going and to see the results of its own moves. That's the quiet power of the agent shape.

The honest part

Where agents fall short.

This is the section the marketing leaves out, and it's the one that saves you money. The same freedom that makes an agent flexible is exactly what makes it risky, and you should walk in clear-eyed:

• Freedom cuts both ways. An agent that can choose its own steps can choose a wrong one — and then build the next step on top of that mistake. Give it real power without guardrails and a small error can compound.
• More steps, more places to drift. A loop with twenty turns has twenty chances to wander off the goal. The longer the chain, the more a good agent design leans on checks and limits, not blind trust.
• Irreversible actions are the danger zone. Reading data is safe to let an agent roam. Sending money, emailing a client, deleting a record — those want a human approving the trigger, not the model firing on its own.
• It is not magic, and it is not free. Every loop is real compute you pay for, and an over-eager agent can run up steps (and cost) chasing a goal a simple script would have nailed in one shot.

The takeaway isn't "don't use agents." It's that a good agent is fenced — pointed at the right job, handed only the tools it needs, and gated by a human on anything it can't take back. The skill is in the fencing, not the freedom.

In the real world

What agents are actually doing out there.

Stripped of the hype, the jobs that suit an agent share a fingerprint: multi-step, a little unpredictable, and made of actions software can take through tools. A few that earn their keep:

• Research and pull-together. Search several sources, read what comes back, and assemble it into one answer — deciding where to look next based on what it already found.
• Inbox and request triage. Read an incoming message, figure out what it's asking, look up the relevant record, and draft a reply or route it to the right place.
• Data wrangling across systems. Take a record from one tool, check it against another, flag the mismatches — the dull cross-referencing that eats an afternoon by hand.
• Monitor-and-act watchdogs. Keep checking a source on a schedule, and when something new shows up that matters, take the first step on it — pull it, log it, notify a person.

Notice what's not on that list: one-step, do-the-same-thing-every-time jobs. "Email this PDF to that address on the first of the month" doesn't need an agent. It needs a three-line script. Reaching for an agent there is using a finish carpenter to drive one nail.

The decision

So — do you actually need one?

Here's the test, and it's quick. Walk through the task in your head and ask: do the steps change depending on what you find along the way?

If the answer is no — same inputs, same steps, same output, every time — you don't want an agent. You want plain automation: cheaper, faster, more predictable, and it won't surprise you. Most "we need AI" jobs are secretly this, and recognizing it saves you real money.

If the answer is yes — the path genuinely bends based on what shows up, the task is several steps deep, and those steps are things software can do through tools — that's the agent-shaped work. And even then, the right build is a fenced one: a clear goal, only the tools the job needs, and a human approving anything the agent can't undo.

That judgment call — agent, simple automation, or leave it to a person — is the whole game, and it's the first thing worth getting right before a line of anything gets built.

Where to take this next

The honest version of "should you build one."

Related reading on the site

If this landed, the next pieces.