Did It Work?
Think about when you are watering a plant. You pour some water. Then what? You look to see if the water soaked in. You check if the soil is wet enough. If it looks too dry still, you pour a little more. You did not just pour water and walk away forever. You checked whether it worked! AI agents have to do the same thing. After they take an action, they must check the result. Did the action do what it was supposed to do? Is the job done? Or does the agent need to try again?
Checking Back After Acting
After the agent acts, the sense-think-act loop immediately starts over. The agent goes right back to sensing. But now the agent is sensing with a specific question in mind: did my action work? Here is a robot example. A robot is asked to push a box from one side of the room to the other. It senses where the box is, thinks about how to push it, and acts by rolling forward and nudging the box. Then the loop restarts. The agent senses again: where is the box now? Is it in the right place? If the box moved to the right spot — great! The job is done. If the box barely moved, or moved the wrong way — the action did not fully work. The agent needs to act again.
After every action, the agent checks: did it work? It senses the new state of the world and compares it to what was wanted. This checking step is what allows the agent to correct mistakes and keep making progress toward its goal.
Checking works the same way for digital agents. Imagine an agent whose job is to book a table at a restaurant. It clicks the Reserve button on the restaurant's website. Then it senses again: is there a confirmation message on screen? Did an email arrive? If yes — the booking worked! Job done. If no — maybe the reservation system was busy, or the table was already taken. The agent senses the error message, thinks about what to try next, and acts again: maybe it tries a different time, or a different restaurant. The checking step after every action is what keeps the agent on track.
Flashcards — click each card to reveal the answer
Fill in the missing word.
Sometimes checking is simple. A thermostat turns on the heater. Then it senses the temperature. If the temperature is rising, the heater is working. Easy check! Sometimes checking is harder. A navigation agent tells a car to turn left. But there might be road construction blocking that turn. The agent senses the obstacle and realizes the turn did not work as expected. Now it needs to think of a different route. The more complex the world, the more important it is to check after every action. A world that surprises you requires an agent that keeps checking.
When an agent checks and finds its action did not work, that is not failure — that is the system working correctly! Noticing that something did not work is the first step to fixing it. An agent that never checks cannot learn from its mistakes.
A robot pushes a box toward a shelf. After pushing, the robot senses that the box is only halfway there. What should the agent do?
A digital agent tries to send an email but gets an error message saying the email address was wrong. What does the agent sense that tells it the action failed?
Try, Check, and Notice
- You will practice checking after acting, just like an agent.
- Get a ball and a target — a laundry basket, a bucket, or a taped circle on the floor works great.
- Stand a medium distance away. Try to toss the ball into the target.
- After EACH toss, check: did it land in the target? How close was it? Was it long, short, left, or right?
- Write down the result of each check: hit, close, or missed.
- Do ten tosses total.
- At the end, look at your check results. Did checking after each toss help you notice how to adjust? Talk about how this is like an agent checking after each action.