Introduction to AI agents

A customer writes, “My delivery has not arrived, and I need it for an event.” A useful response requires more than a friendly paragraph. Someone must understand the problem, find the order, check the latest delivery information and decide which permitted options to offer. An AI agent is a software system that uses a model to choose steps towards a goal, with information and actions made available by its application.

The word agent does not mean a digital person or an employee without supervision. It describes how software works. The system can choose a next step rather than follow only a fixed sequence, but its competence and authority still depend on its design. A support agent might investigate a late delivery while remaining unable to issue a refund.

Imagine a new employee #

Imagine welcoming a new colleague on their first morning. They are good at reading and writing, but they do not know your company. You give them a handbook, explain their responsibilities, show them the customer service screen and provide a phone. You also tell them which decisions require a supervisor.

A language model supplies part of the reading and writing ability. Instructions act like the job description. Business documents act like the handbook. Tools provide controlled access to systems. Rules and human review establish the limits. Without these additions, even a capable model is like a new colleague answering customers from general experience rather than company evidence.

The analogy has limits. A person can notice physical events, develop relationships and accept responsibility in ways software cannot. An agent sees only information its application supplies. It does not automatically know that a parcel arrived, that a policy changed or that a customer is authorised to access a record. Those facts need reliable sources.

Chatbot, script or agent? #

A chatbot is a conversational interface: something you can message and receive a reply from. It may follow fixed rules, use a language model or front an agent. A script is a program that follows steps written in advance. An agent uses a model to choose some steps as the situation unfolds. These categories overlap rather than form a ladder of quality.

A script can send an e-mail and an agent can produce only text. Taking an action is therefore not the only distinction. The important question is where decisions about the next step come from. For a precisely defined task such as adding invoice totals, ordinary software may be simpler and more reliable than asking a model to decide what to do.

A chatbot can also be useful without becoming an agent. If visitors need opening hours and directions, a clear answer may be enough. Adding tools and a decision loop introduces more things to test. Start from the work the user needs done, not from a desire to attach the agent label to every conversation.

The perceive, decide and act loop #

Perceive means receive information: a message, a document or a tool result. Decide means choose a next step using the goal, instructions and available evidence. Act means request an allowed operation or produce a response. The application returns the outcome, allowing another decision. This repeated pattern is called a loop.

For the late delivery, the first decision may be to ask which order the customer means. Once the application identifies an authorised order, the agent can request its status. If tracking reports a delay, the agent can explain the available options. If the lookup fails, it should not pretend to have checked successfully. The new information changes the next step.

The loop needs a stopping rule. “Keep trying until the customer is happy” is too open-ended: a system could repeat failed calls or make increasingly unsupported promises. Better stopping conditions include receiving a confirmed result, encountering an unavailable service or reaching a decision reserved for a person. The application should enforce limits on work as well as on permissions.

Where agents help #

Good starting tasks involve varied language but a recognisable business process. Customers describe the same problem in many ways. Employees ask questions without knowing a document’s title. A model can help interpret these requests while ordinary software remains responsible for reliable record access and changes.

Notice that these examples have a clear finish: an answered question, a completed set of requirements or a prepared ticket. They do not ask the agent to “run the department.” Smaller responsibilities are easier to explain to users, evaluate against examples and hand over when something goes wrong.

What agents are not good at #

Agents can misunderstand ambiguous requests, miss relevant evidence and confidently state something false. They are especially risky when asked to make exact calculations without a calculation tool, infer private facts they have not received or handle long chains of dependent decisions without checks. A fluent explanation does not prove that a task was completed correctly.

Some decisions also involve responsibility that cannot sensibly be delegated to a text-generating system. Approving an exceptional payment, interpreting a serious complaint or making a consequential employment decision may require accountable human judgement. An agent can organise evidence for the person without making the final decision itself.

Before choosing an agent, ask what a wrong answer could change. A weak draft can be edited; an unauthorised payment may be difficult to reverse. Start with assistance that a person can inspect, then expand authority only where evidence, permissions and evaluation justify it. More independence is a design choice, not an automatic measure of progress.

Related: On AIVAX, the configured runtime that brings a model, instructions and capabilities together is called an AI gateway.

What’s next: Meet the engine at the centre of this system in What is an LLM.