Module 1 · Foundations of Agentic AI · scripted
Shadows on the Wall
Everything here is made of words
Stop and notice what every lesson so far has in common. The conversation, the trajectory, the tools, the parameters, the protocol — all of it is words. The LLM's entire understanding of any situation is built from words in our language. It does not see the kitchen, feel the batter, or watch the ferry leave. It reads.
There is a very old story about minds in exactly this position.
The allegory of the cave
Plato's cave, in brief
Prisoners sit chained in a cave, facing a wall. They have never seen anything else.
Behind them burns a fire. Between the fire and the prisoners, things are carried past — people, animals, objects. The fire casts their shadows onto the wall.
The prisoners see only the shadows. Their entire understanding of the world — what things are, how they move, what follows what — is built from shadows. Not the things themselves. Representations.
And from those representations, they build a model: they learn to predict which shadow comes next.
The cave
The shadows are not the world. But they are cast by the world — and a careful watcher can learn an astonishing amount of the world's structure from them.
The LLM in the cave
When an LLM is trained, it learns from our descriptions of things — our word-concepts and the relationships between them. Not by observing water, gravity, kitchens, or ferries, and never by experiencing them. Our descriptions are the shadows on its wall.
What is remarkable is how good the model it builds from those shadows is — and that it can cast shadows back: given a description of a situation, it projects words describing what it thinks happens next.
A full glass of water sits at the very edge of a table. A toddler runs past and clips the tablecloth. What happens in the next three seconds?
The glass tips over the edge and falls. Water spreads across the floor faster than anyone can react; the glass shatters or bounces depending on the material. The toddler either keeps running, oblivious, or stops at the sound. Someone nearby says something sharp, and the immediate job becomes glass first, water second, toddler reassured.
We stand in an interesting symmetry with it. We learned our model of the world directly — hands in the batter, eyes on the fjord. The LLM learned its model from our descriptions. And yet both of us can, in many cases, accurately predict what will happen, and both of us project our predictions outward as words.
The context is the shadows we cast
Now the practical point, and it is the most important sentence in this lesson: when you converse with an LLM, you are the fire. The conversation is the wall. What you write is the shadow the world casts.
If we cast the shadows badly — vague, incomplete, missing what matters — the model's ability to choose actions and steer toward the state we want is poor. Not because its model of the world is poor, but because our projection of the situation into words was:
My study's results look weird. What statistical test should I use to check my two groups are different?
For comparing two groups, the standard choice is an independent-samples t-test — check normality first, and if it fails, use a Mann-Whitney U test instead.
My study's results look weird. Details: two groups of nurses (n=9 and n=11) rated workload on a 1–5 scale after each shift; ratings are ordinal, heavily skewed toward 4–5, and each nurse contributed ~30 shifts, so observations within a nurse are not independent. What test?
Then a t-test is wrong twice over: the data are ordinal and the observations are clustered within nurses. You want a method that respects both — for example an ordinal mixed-effects model with a random effect per nurse. And "weird" may just be the ceiling effect: with most ratings at 4–5, differences compress near the top of the scale.
- In Example 2, the model assumed things the user never said. What — and why those?
- What, exactly, did the words in Example 3 preserve about the world that Example 2's words destroyed?
Engineering agents is engineering shadows
There are countless ways to describe any situation, and every description keeps some of the world's structure and throws the rest away. That is the real work ahead of us. An agent cannot directly observe anything: its goal is a shadow, its tool descriptions are shadows, every result a tool sends back is a shadow. So a key part of engineering an AI agent is exactly this: taking some enormously complex aspect of the world and projecting it into words — casting shadows careful enough that a mind which sees only the wall can still act well in the world behind it.
Take one complex situation from your research — a process, a dataset, an organization. Write two projections of it: one in three sentences, one in ten. Give each to an LLM with the same question, and compare what it got right and wrong under each shadow. What did your longer projection preserve that mattered? What did both throw away?