A chat assistant answers. An AI agent acts. Ask an assistant how to reschedule a dentist appointment and it explains the steps. Ask an agent and it checks your calendar, contacts the office's booking system, picks a time that works, and confirms. That shift, from producing text to taking actions in the world, is why agents are both useful and in need of careful design.
This is the third step in our AI and agents learning path. It builds on How AI Language Models Work and Using AI Well.
The parts of an agent
Most agents combine the same building blocks:
- A model that reads the situation and decides what to do next. This is the language model from the first step in this path.
- Instructions that define the agent's job, its rules, and what it must never do.
- Tools the agent can use: searching a knowledge base, reading a calendar, looking up an order, sending an email, updating a record. Each tool is a specific, limited capability the builders choose to provide.
- Memory of the task so far, and sometimes of past interactions, stored deliberately by the application.
- Guardrails that check inputs and actions, block what is not allowed, and route certain decisions to a person.
Our interactive agent anatomy lab walks through these layers one at a time.
The loop: plan, act, check
Where an assistant answers once, an agent works in a loop:
- Understand the goal. "Reschedule my Thursday appointment to next week."
- Plan a step. "First, find the existing appointment."
- Use a tool. Look up the calendar.
- Check the result. "Found it: Thursday at 3 p.m. with the dentist."
- Repeat until the goal is met: find open times, choose one that fits, book it, cancel the old one.
- Report back, or stop and ask when something is unclear.
Each turn of the loop is a chance to make progress, and also a chance to make a mistake. A wrong step early on can carry through every step after it, which is why checking results at each step matters.
What agents are good for
Agents fit work that is repetitive, follows rules, and needs a little judgment to handle variety. Examples include:
- Answering customer questions using a company's own policies and order data, and handing anything unusual to a person.
- Taking phone calls after hours, booking appointments, and capturing details for staff in the morning.
- Reading incoming documents, pulling out the key fields, and entering them into a system.
- Gathering information from several internal systems and drafting a report for a person to review.
Why guardrails matter more for agents
When an assistant makes a mistake, you get a wrong answer that you can ignore. When an agent makes a mistake, it may send the wrong email, change the wrong record, or spend money. So well-built agents follow a few principles:
- Least privilege. Give the agent only the tools and permissions the job requires. An agent that answers billing questions should be able to read invoices, not issue refunds of any size.
- Human approval for consequential actions. Payments, deletions, legal commitments, and anything hard to undo go to a person first.
- Clear boundaries. The agent knows what it must decline or hand off, and does so reliably.
- Logging. Every action is recorded so people can see what happened and why.
- Testing before launch. Agents are run against a large set of realistic scenarios with known correct outcomes, including attempts to trick them, before they meet real customers. The same set is rerun whenever the model, the instructions, or the tools change.
That last practice is ordinary software testing discipline applied to a new kind of system, and it is where the testing and AI learning paths meet.
When not to use an agent
Not every automation needs an agent. If a task always follows the same steps in the same order, a fixed workflow (step one, then step two, then step three) is cheaper, faster, and easier to test. Agents earn their cost when the inputs vary enough that rigid rules break down. A useful question before building anything: "Could we draw this as a flowchart with no 'it depends' boxes?" If yes, start with a workflow. Our article Workflow or Agent? covers this decision in depth.
Try it yourself
Pick a task you do regularly that involves several steps, such as planning a group study session or ordering supplies for a club. Write it down as an agent would see it:
- What is the goal?
- What tools would the agent need (calendar, messaging, a spreadsheet, a store)?
- What are the steps in the loop?
- Which actions should require your approval before the agent proceeds?
- Could this be a simple workflow instead? Why or why not?
The next level of this path moves to how organizations choose between workflows and agents, pick the right model for each task, and control the cost.
Look up any unfamiliar term in our glossary.