SLIDE 6: Section Intro, LLM Foundations for Agents

Section 2: LLM Foundations for Agents

Agents run on LLMs. Every agent you build, every tool call it makes, every reasoning step it takes is powered by an LLM generating tokens. If you do not understand how LLMs work, where they fail, and what knobs you can turn, you will build unreliable agents and you will not be able to debug them when they break.

This section covers LLMs at interview depth, not research depth. You need to explain transformer architecture, tokenization, and inference clearly in 2-3 minutes during an interview. You do not need to derive the attention equation or explain backpropagation.

What We’ll Cover

  1. Transformer architecture: what it does, how to explain it in an interview
  2. Tokens and context windows: the agent’s working memory limit
  3. Inference, temperature, and determinism: why agents need predictable outputs
  4. LLM limitations that directly drive agent architecture decisions
  5. Model selection: choosing the right model for your agent’s tasks

Where This Shows Up in Agent Systems

Agent Component LLM Foundation It Depends On
Core reasoning loop (Section 3) Transformer inference, next-token prediction
Tool calling (Section 4) Structured output generation, low temperature
RAG and retrieval (Section 6) Context window limits, token budgets
Memory systems (Section 5) Context window overflow, conversation history
Cost and latency (Section 11) Token pricing, model selection, inference speed
Reliability (Section 8) Hallucination, non-determinism, failure modes

Every section in this course builds on these foundations. When something goes wrong with your agent, the root cause often traces back to one of the LLM behaviors covered here.

SLIDE 7: Transformer Architecture at Interview Depth

What a Transformer Does

A transformer takes a sequence of tokens as input and predicts the next token. It does this repeatedly, generating one token at a time, until it produces a complete response. That is the entire mechanism behind ChatGPT, Claude, Gemini, and every other LLM powering today’s agents.

The Three Concepts You Need