What is an AI agent, and how is it different from an LLM?

An AI agent is a system that uses a language model to decide what to do, and then actually does it. That last part is the whole difference. Ask ChatGPT about an overdue invoice and you get a well written explanation of how to chase it. An agent opens your accounting software, finds the invoice, and sends the reminder.

Most conversation about AI still treats the model as the entire system. Prompt goes in, answer comes out. That is one component of something larger.

An LLM answers. An agent acts.

A language model can tell you how to update a customer record. That is not the same as updating it. It can draft an email. That is not the same as sending one. It can identify a customer by name. That is not the same as pulling their actual order history out of your database.

The model supplies the reasoning. Everything around it determines what that reasoning can reach.

A rough shape for what an agent does: input arrives, the model works out what is being asked and what information it needs, it picks a tool, the tool returns something, the model checks whether that something is what it expected. If yes, it acts. If no, it retries, asks a person, or stops.

Reading that back, the interesting parts are not the model. They are the checking and the stopping.

What an agent needs beyond the model

An agent doing real work needs several things the model does not provide on its own.

Context and instructions, so it knows what it is being asked to do and what counts as done.

Tools, so it can reach systems outside itself. Reading your invoices requires a connection to wherever your invoices live.

Memory and state, so a workflow that runs every Monday knows what happened last Monday.

Permissions, because read access is not write access, and write access is not permission to delete. An agent that can issue refunds should be a deliberate decision rather than a side effect of connecting a payment tool.

Verification, so that "the workflow ran" is not confused with "the task succeeded."

Observability, so a person can see what the agent looked at, what it decided, and what it did.

Escalation, for the situations where the system should not make the final call.

None of these are exciting in a thirty second demo. They are what separates a demo from something you would let near your customers.

The failure path is part of the product

The easiest workflow to build is the one where everything goes right. The user supplies exactly the information you expected, the tool responds the way it always does, the model makes the correct call.

Real workflows are less cooperative. Information goes missing. A service is down. The model is not confident. A step needs a human to approve it. Sometimes a workflow completes successfully and still fails to achieve the thing it was supposed to achieve.

An agent worth relying on treats those as paths rather than dead ends. Something unexpected happens, the agent notices, and then it retries, asks, escalates, or stops.

That behaviour is most of the difference between an impressive demo and something a business can put in front of customers.

What this looks like in practice

Here is a run from an invoice chaser built on Hivework. It reads last week's orders, finds the ones nobody paid, and sends reminders.

It sent 13 and held 1. One invoice had something about it the agent was not confident enough to act on, so it stopped and asked. The person answered from their phone, over iMessage, without logging into anything.

Every step has a cost next to it. That is the observability part made concrete. You can see what the agent did and what each decision cost, rather than trusting that it worked. The whole run cost 2.4 cents. Not because cheapness is the point, but because a receipt you can read is a different thing from a black box that either worked or did not.

Hivework agents work in the accounts you already have, on a schedule you set, and stop for a person when they should. Every run comes back as a receipt like this one. A run that fails refunds itself.

Step	Cost
Read 14 new orders	0.4c
Matched them to customers	0.9c
Sent 13, held 1	1.1c
Whole run	2.4c