IBM's Maya Murad gives one of the clearest short explanations of what an agent actually is. It moves quickly and gets a little technical in the middle (she mentions "compound AI systems" and "ReAct"). Don't worry about the jargon. Watch it for the shape of the idea, and the few minutes below will make the core stick.
What this means for you
Start with what a plain AI model can and can't do. On its own, a model only knows what it was trained on. Maya's example: ask it "how many holiday days do I have left" and it can't answer, because it has never seen your HR system. Great for drafting an email or summarising a document, but it can't reach into your world.
The first step up is to build a small system around the model: give it access to your data so it can look up your holiday balance and answer properly.
An agent goes one step further again. Instead of you spelling out every step, you put the model in charge of working out the steps itself. You give it a goal, and it makes a plan, takes actions to get there, and keeps going until it's done. Maya splits that into three things an agent can do, and they're worth holding onto, because the rest of this course follows them:
- Reason: break a goal into steps and think them through, rather than blurting the first answer.
- Act: use tools to get things done, search, look something up in a system, run a calculation, send something.
- Remember: hold on to what it has learned so it does better next time.
Her own example makes it real: "how much sunscreen should I take to Florida next month?" That's not one question, it's several: how many days am I going, how sunny will it be, what's the recommended amount, then the maths. A plain model guesses. An agent works through each part and comes back with a real answer.
One honest point she makes, and it's an important one: not everything needs an agent. If a task is simple and always the same, you're better off just setting fixed rules. Agents are worth it on the messier, multi-step work. And in almost every case, there's still a person checking the result.
Take month-end in a finance team. Someone spends the best part of two days pulling figures out of three systems and reconciling them into one report. That's a goal with clear steps and a handful of lookups, the exact shape an agent suits. The agent does the stitching; the person keeps the judgement and gets those two days back to spend on it.
Try this
Think of one job on your team that's really several steps stitched together: onboarding a new starter, booking and expensing a trip, chasing a sign-off across three people. That shape, a goal with a few steps and lookups along the way, is what agents are built for. Hold one in mind as we go.
Common questions about AI agents
What is an AI agent in simple terms?
An AI agent is a model that you put in charge of a goal rather than a single question. Instead of you spelling out every step, the agent works out the steps itself: it makes a plan, takes actions to reach the goal, and keeps going until it's done. A useful way to remember it is that an agent can do three things a plain model mostly can't:
- Reason: work out a plan, rather than blurting the first answer.
- Act: take the steps to reach the goal.
- Remember: hold on to what it has learned.
For your team, that means an agent suits a goal, not a one-line query. Onboarding a new starter is a goal with several steps, so it fits an agent. Checking one fact does not, and a plain model is the simpler choice there.
What is the difference between an AI agent and a chatbot?
A plain model or chatbot only knows what it was trained on, so it can't tell you how many holiday days you have left because it has never seen your HR system. It is good inside that limit, drafting an email or summarising a document. An agent adds the ability to look things up in your systems, take actions, and chain the steps together to reach a goal. So when a job is really several steps stitched together, that's where an agent fits and a chatbot doesn't. Asking your holiday balance and then booking the time off is two steps in your systems, so it suits an agent. A single self-contained ask is enough for a chatbot.
Can an AI agent do things on its own without a person?
Up to a point. An agent can plan and take the steps itself, so you give it the goal rather than spelling out each action. But in almost every case there's still a person checking the result. Take finance month-end: the agent does the stitching across systems and reconciles the figures, while the person keeps the judgement and reviews the result.
Note: treat agents as doing the legwork, not removing the human check.
When should you not use an AI agent instead of a chatbot or fixed rule?
Not everything needs an agent. If a task is simple and always the same, you're better off just setting fixed rules. Agents earn their place on the messier, multi-step work: a goal with a few steps and lookups along the way, like onboarding a new starter or chasing a sign-off across three people. A task that runs the same way every time is better left to a fixed rule.
Note: before reaching for an agent, ask whether the job is genuinely multi-step or just a fixed routine.