What is an AI agent? A plain answer for people who run a business
An AI agent is a language model using tools in a loop, within limits you set. What that means in practice, what to trust it with, and three agents we run.
- An AI agent = a language model + tools + a loop + limits.
- The model decides the next step, a tool carries it out, the result goes back to the model, and the loop repeats until the job is done.
- What makes an agent useful is the tools. What makes it safe is the limits.
- Trust agents freely with work that can be undone. Put a person in front of anything that cannot.
An AI agent is a language model that can take actions, not only write text. You give it a goal, a set of tools it may use and some limits. It then works in a loop: decide the next step, use a tool, look at the result, decide again, until the goal is reached or a limit stops it. A chat assistant that only answers is not an agent. The same assistant becomes one when it can search your files, query a database or save a draft in your mailbox, and chooses for itself when to do so.
That is the whole definition. The rest of this article is about what it means for someone who runs a business, and the three agents we run every day.
The four parts
1. A model: the part that decides
A large language model, such as Claude, GPT or Gemini, reads the goal, the conversation so far and the results of earlier steps, and proposes what to do next. It is good at reading messy input and at choosing between options. It does not remember anything between runs unless you give it memory, and it can be confidently wrong.
2. Tools: the part that acts
A tool is a named action with clear inputs: search_mail(query), create_draft(to, subject, body), list_cases(status). The model asks to use a tool, ordinary software carries it out, and the result goes back to the model. The tools define what the agent can affect in the world. An agent with no tools is a chatbot. An agent with a "delete everything" tool is a risk.
3. A loop: the part that persists
An agent does not stop after one answer. It keeps going, step by step, and adjusts when a step fails. This is what lets it finish a task with ten steps, and also what lets it go wrong in ten steps if nothing checks it.
4. Limits: the part that keeps it useful
Limits are which tools it has, what each tool is allowed to touch, how many steps or how much money a run may use, and which actions need a person. Most of the work of building a good agent is here, not in the prompt.
What you can trust an agent with
The most useful test we know is the one in our Academy lesson Agents at full autonomy: can the action be undone?
- Undoable: reading, summarising, sorting, labelling, drafting, filing, opening a ticket, writing to a work branch. Let the agent do these without asking. Asking permission for each one only trains people to click "yes" without reading.
- Not undoable: sending to a customer, paying, deleting, publishing, changing access, changing a database schema. These wait for a person, and the agent should not even have a tool that can do them alone.
There is a second test: does the agent know when it is stuck? A well-built agent that meets a task it has no recipe for does not improvise. It looks for an existing procedure and, if there is none, asks for one. We call that the escalation reflex. An agent without it will confidently invent a way through, which is exactly where the damage happens.
Three agents we run
The sales drafting agent
It reads incoming sales enquiries that carry a label, loads the company's knowledge base and writes a reply as a draft in the mailbox. Its tools can read mail, create drafts and record a case. It has no tool that sends email. A person approves every reply on a small board. Details are in AI inbox management.
The image agent behind this blog
The illustration above was made by an agent calling an n8n webhook, which asks an image model for a picture and returns a link. A second webhook commits the file to a work branch in git, never to the live site. The agent never sees the image service's API key. The key stays inside n8n. The agent has a tool, not a secret.
The coworkers that keep our repositories
Several AI agents work on our websites and tools at the same time. Each picks tickets from a shared board, works on its own branch or files, and leaves notes for the others on a small message bus in a database rather than through a person. They act on their own for anything reversible. A short, written list of gates is reserved for a human: merging someone else's branch, new infrastructure, deleting, database schema, DNS, spending money and access rights. The list is short on purpose. Short lists get followed.
What an AI agent is not
- Not a person. It does not carry context from last week unless you store it, and it does not feel responsible. Responsibility stays with whoever set the limits.
- Not a fixed workflow. A workflow does the same steps every time. An agent chooses its steps. Most good systems combine both: a predictable workflow with one agent step where judgement is needed.
- Not magic at scale. Every step costs a model call. A loop with no step limit can be slow and expensive. Models, limits and cost shows how to keep that in check.
Where to start
Use an agent on something you already do by hand, where a mistake is cheap and visible: sorting a folder, drafting replies you will read anyway, summarising a long thread. Watch what it does with the tools you gave it. Widen its limits only when you have seen it behave. The Academy lesson Working with an AI agent teaches the everyday loop of describing, checking and refining, and the glossary explains the vocabulary in plain words. For the jobs worth handing over first in a small company, see AI agents for small business.
Frequently asked questions
What is the difference between a chatbot and an AI agent?
A chatbot answers in text and stops. An agent can also act: it calls tools such as a search, a database query or a draft in your mailbox, looks at the result and decides the next step, until the task is done or it hits a limit.
Is an AI agent the same as automation?
No. A classic automation follows fixed steps you designed. An agent chooses its steps at run time. In practice the two are combined: a fixed workflow does the predictable parts and calls an agent for the step that needs judgement.
Are AI agents safe to use in a business?
They are as safe as the limits around them. Give an agent only the tools it needs, let it act freely only on what can be undone, and require a person for anything irreversible such as sending, paying or deleting.
Do I need to code to use an AI agent?
Not to use one. Assistants like Claude or ChatGPT already work as agents when you connect tools. Building your own agent into a business process is easier with a workflow tool such as n8n, which needs little or no code.
Written from our own production systems. Drafted with AI assistance and reviewed by Kamil Hrbáček.
Want this running in your company?
Flowsmith designs and operates automations like this for small and mid-sized teams — with a person approving anything that leaves the building.
