PYTHON · NUMPY

What is an agent?

Draw the loop that makes something an agent, and find the one place it can touch its world.

01Challenge

Sweep how hard an agent couples to its world and find the optimum — it is neither zero nor infinity.

02Model

An agent is a loop: sense, decide, act, and the world answers. Software AI sits far from the place where that loop closes. Physical and embodied AI sit right on it — a wheel on gravel, a gripper on glass, an antenna in the air. That contact is an interface, and an interface can be matched well or badly. This module makes the point in the smallest possible way: an agent that barely couples to its world cannot move it, and an agent that couples too hard fights its own delay and rings. Somewhere between is a best setting. Finding that setting, in every domain intelligence touches, is what the rest of this course is about.

03Guided practice

1. The loop. The agent senses the world late, decides, and acts; the world answers, and the answer arrives late too. 2. The knob. Gain is how hard it insists. 3. The sweep. Track the error as gain rises. It falls, bottoms out, then the loop destabilises — the same shape you will meet as reflection, as capacity, as contact chatter, and finally as surprise.

04Feedback

PASS WHEN PASS when the best gain is interior — neither the weakest nor the strongest.

If your run fails, check:

A cell errored — run it again.
Expected the tracking error to bottom out at an interior gain.

05Retrieve & space

What are the four steps of a sensorimotor loop? (sense, decide, act, and the world answers).
Why does too much gain hurt? (it fights the delay in its own loop and rings).
Where does an agent actually touch its world? (at its interfaces — its sensors and its actuators).

06Mastery & project

You can diagram an agent's loop and name where it couples to its world.

The habit: find the interface first; that is where the physics of intelligence lives.