This is the architecture video. If the first lesson answered "what's an agent?" this one shows you the engine. You'll see how information flows in, gets processed, becomes a decision, and turns into action. There's a worked example about booking a trip that ties it together. Worth watching for the shape of the thing.
What this means for you
Every AI agent breaks down into three layers: sensing, thinking, and acting.
Sensing is how the agent gets information in. That might be text you type into a chatbot, data from a camera or microphone, or an API call that flags something has happened. It's the agent's perception of the world.
Thinking is where the real work happens. The agent takes what it sensed and processes it against a knowledge base. That knowledge base holds three kinds of thing: facts your organisation holds (prices, locations, availability), rules and policies the agent must follow (spend caps on hotels, preferred suppliers), and the goals you've asked it to achieve. The agent reasons through all of this, breaks down complex tasks into smaller steps, and works out a plan. It uses simple if-then logic, pattern recognition from examples it has seen before, and language model reasoning to think itself through to an answer.
Acting is the output. The agent might generate text or speech, write to a database, trigger an alert, or control something physical. In the travel booking example, the action is talking to the airline and hotel systems to make the reservations.
The feedback loop closes the circle. After the agent acts, you or the system itself checks: did that work? Did it match what we wanted? The agent learns from that feedback, either because you give it a thumbs up or down, or because it checks itself whether an action got it closer to the goal. Over time, it gets better and more personalised.
Picture a support team handling refund requests. Each request needs approval from three departments, a check against policy limits, and sometimes a call to the finance system. That is exactly the work an agent can handle: it takes in the request details (sensing), checks policy rules and spending caps against a knowledge base (thinking), approves or escalates the decision and logs it (acting), and learns whether its judgement was right when a manager reviews it next week (feedback). The team doesn't disappear. Someone still checks tricky cases. But the routine ones flow faster, and the team spends their time on what really needs a person.
Try this
Find one process your team runs where the same policy rules apply every time, data lives in a system or document, and a person has to look things up and make a decision. That is probably a job that fits the sensing-thinking-acting anatomy. Write down the three layers for it: what information comes in, what knowledge or rules matter, and what action does the person take. You have just sketched out an agent.
Common questions about how AI agents work
How does an AI agent work?
An AI agent works by moving information through three layers: sensing, thinking, and acting. Sensing is when it takes information coming in, such as text you type or an API call flagging that something has happened. Thinking is when it processes that against a knowledge base of facts, rules, and goals to work out a plan. Acting is when it produces an output, like generating text, writing to a database, or triggering an alert. A feedback loop closes the circle, so the agent checks whether the action worked and improves over time.
Take booking a work trip. The agent senses the dates and destination you give it, thinks by checking those against your stored preferences and spending caps, and acts by booking with the airline and hotel systems.
What are the main parts of an AI agent?
An AI agent has three core parts: sensing, thinking, and acting.
- Sensing: how it gets information in.
- Thinking: where it reasons over a knowledge base of your facts, your rules and policies, and the goals you set.
- Acting: the output it produces.
A feedback loop sits around these three, letting the agent check its work and get better over time. On a customer support desk, sensing is the incoming ticket, thinking is checking it against your refund policy, and acting is approving the refund or escalating it.
How do AI agents sense, think and act?
Sensing, thinking and acting is the basic anatomy of an AI agent. Sensing is when it takes information coming in. Thinking is when it processes that against a knowledge base of facts, rules and goals. Acting is when it produces an output, such as making a booking or logging a decision. A feedback loop then checks whether the action matched what you wanted, either through a thumbs up or down from you or the agent checking itself, so it improves and becomes more personalised over time.
On a support desk handling refund requests, the agent senses the request, thinks by checking it against your policy limits, and acts by approving the refund or escalating it to a manager.
Where does an AI agent get its knowledge to make decisions?
An AI agent reasons against a knowledge base that holds three kinds of thing. The first is facts your organisation holds, such as prices, locations and availability. The second is rules and policies it must follow, such as spend caps and preferred suppliers. The third is goals you have asked it to achieve. It works through all of this using simple if-then logic, pattern recognition from examples it has seen, and language model reasoning. For a work trip, that knowledge base is the prices, locations and availability it looks up, the business spend caps and preferred travel partner it must follow, and the goal of booking the trip, which the agent reasons over to make the reservations.
Note: an agent is only as good as the facts, rules, and goals you give it, so the routine processes where your policies are clear and written down are the best places to start.