Foundations · Chapter 2 of 2

How models think

Tokens, context windows, and why agents sometimes get weird.

You don't need a machine learning degree to build with AI. You need a working mental model for why an agent can be brilliant for an hour and then suddenly confused. Four ideas explain almost all of it.

Tokens

Models read and write in chunks of roughly 3 to 4 characters called tokens. Everything you send, from prompts to files to pasted errors, is measured and billed in them.

Context window

The model's working memory: everything it can consider at once. Your instructions, the files it read, the whole conversation, all of it has to fit.

Hallucination

Models predict plausible text; they don't look facts up. Sometimes they confidently invent a function or a setting that doesn't exist.

Agents

An agent is a model plus tools plus a loop: it can read files, run commands, see what happened, and try again, without you pasting anything.

What this means in practice

Long chats degrade. As a conversation fills the context window, older details get fuzzy and quality drops, usually well before the window is technically full. The fix is cheap: one chat per task, and start fresh when things get weird. A clean description in a new chat beats fighting a foggy one.

Run it and see. Because models predict rather than know, treat every answer as a draft that must run, not a fact to believe. In coding this is a superpower rather than a danger: wrong code reveals itself the moment it executes, and the agent reads its own errors and fixes them.

The harness matters as much as the brain. A raw model can only write text. What makes modern tools feel magical is the agent loop wrapped around the model, which is why an agent running a mid-tier model routinely beats copy-pasting into the smartest chat window on earth. When you compare tools, you're really comparing harnesses.

The one habit to build now

Ask to see it working, not just to hear it's done. “Run it and show me” is the highest-value sentence in vibe coding.

← PreviousAI models Next →Opening the terminal

AI Clubhouse · Vibe Coding Guide