How AI agents work

01Agent versus chatbot

A chatbot produces text in response to text. An agent can also act: look up an order, change a booking, issue a refund within a limit, send a confirmation, create a ticket. The difference is tools and a loop. The model is the same kind of system in both cases; what changes is what it is connected to and how the conversation is structured.

02The loop

1. Goal. The agent receives a task — from a customer's message, a schedule, or an event in another system.
2. Reason. The model reads the goal, the context it has been given, and the list of tools it may use, and decides what to do next.
3. Act. It calls a tool: a database query, an API, a search over documents, a message to a person.
4. Observe. The tool's result comes back into the conversation, and the model reads it.
5. Repeat or stop. It continues until the task is complete, it hits a rule (a limit, an approval), or it decides it needs a person.

03The parts of an agent

Model

The language model that reads and decides — Claude or a comparable frontier model, chosen for quality, cost and where data may go.

Instructions

Who the agent is, what it may and may not do, how it should speak, when it must escalate. The most important document in the system.

Tools

Functions the agent may call, each with a description the model reads and a permission boundary the code enforces.

Retrieval

Search over your documents and records so the agent answers from your facts, with citations.

Memory

What it remembers within a task, and — carefully — across tasks: a customer's preferences, an open case.

Guardrails

Limits on actions, checks on outputs, approval steps, and a trace of every decision.

Evaluations

A suite of real scenarios run on every change, so behaviour cannot drift silently.

04Where agents fail

• Too much freedom. An agent with broad tools and vague instructions will eventually do something surprising. Least privilege is the rule.
• No evaluation. A prompt tweak that fixes one case breaks three others, unnoticed.
• Long tasks. Errors compound over many steps; good agents are scoped to short loops with checkpoints.
• Ambiguous goals. If the task has no clear right answer, the agent cannot know when it is done.

05Single agents and multi-agent systems

Most business problems are best served by one well-scoped agent with a clear role. Multi-agent systems — a coordinator delegating to specialists — help when tasks are genuinely separable (research, then draft, then check), but they multiply cost and failure modes. Start with one.

06Deploying an agent safely

1. Give it one role and the minimum tools.
2. Write the evaluation set before the instructions.
3. Put approval steps on anything consequential.
4. Log every decision and tool call.
5. Launch to a slice of traffic, read the traces, widen.