AI agent vs LLM vs model: who does what?
AI agent vs LLM vs model, in short: a model is a computer program that has learned from examples. An LLM, short for large language model, is a model that has learned to read and write text. An AI agent is software built around an LLM that gives it a job, the means to do it and rules to follow. The LLM writes. The agent gets work done.
The three words, one at a time
- Model. The broadest word. A model is a program that nobody wrote rule by rule. It learned its skill from a very large number of examples.
- LLM. A model that learned from a huge amount of text. You give it text and it writes text back. The AI behind ChatGPT, Claude and Gemini is an LLM in each case.
- AI agent. A program that puts an LLM to work. It holds the job description, the list of things the LLM may do, such as looking up a customer or sending an email, and the rules about when a person has to approve first.
A car is a fair comparison. The LLM is the engine. The agent is the rest of the car: steering wheel, brakes and today's route.
What an LLM cannot do on its own
An LLM takes text in and gives text out. On its own it cannot open your customer list, send an email or remember what you discussed last week.
The agent adds these abilities. When an AI "looks something up", the LLM only writes a short request, such as "find the unpaid invoices". The agent does the looking up and hands the result back to the LLM as text.
One request, step by step
Say you run a small trading company and tell your AI: "Remind the customers whose invoices are more than 30 days late."
| Step | What happens | Who does it |
|---|---|---|
| 1 | Your request is put together with the job description and the rules | Agent |
| 2 | The request is read and a first move is chosen: find the late invoices | LLM |
| 3 | The late invoices are looked up. Four come back | Agent |
| 4 | Four polite reminders are written | LLM |
| 5 | The drafts are read and confirmed | You |
| 6 | The emails go out and a note is added to each invoice | Agent |
The LLM appears twice: once to decide and once to write. Everything that touches your data is done by the agent.
What this means when you choose AI software
Because the parts are separate, there are two separate questions to ask.
- Which model is inside? This decides how fast the answers come, how well difficult tasks are handled and what each use costs.
- How is the agent set up? This decides what the job is, what the AI may read, what it may do and when it must ask first.
This is why two products built on the same LLM can behave very differently.
When the result is wrong, check in this order
The model is rarely the first thing to change.
- Could it see the right information? If the AI cannot reach a record, no model can use it.
- Were the instructions clear? One job, described plainly.
- Was it allowed to do what the job needs? If it may not search, it can only guess.
- Only then: is the model too small for the task?
How Worknano handles this
In Worknano the two parts are kept apart on purpose. Each agent has a name, instructions, tools (the actions it may take) and a model, so you can change the model without touching the rest. A fast model suits short checks and a balanced one suits planning with several steps. You can also connect a model of your choice, including open-source models or a model your company runs itself. With regional hosting in the EU, the USA or APAC, the AI processing stays in that region.
- Which AI model does Worknano use?
- AI agents: reusable AI tasks that run on your records
- Build your own agent: instructions, tools and model
Create a free workspace, open an agent and see its model, tools and instructions side by side.