Lesson 13 · Slide 1
Lesson 13 — The LLM & Agent Era
What changed 2022–2026 — LLMs, RAG, agents, MCP, the new product landscape.
IT Landscape
Module 4
High-volatility — as of 2026
Lesson 13 · Slide 2
Where we are in 2026
The most disruptive software shift since the smartphone — happening over three years.
- Nov 2022 — ChatGPT released to the public.
- By 2026 — AI assistants are embedded in IDEs, browsers, the OS, customer-service stacks, and a growing share of search.
- Workflows are being rewritten — coding, writing, support, research, design.
Snapshot — every number, every brand, and every pricing figure in this lesson has a “best before” date.
Lesson 13 · Slide 3
What is an LLM?
A Large Language Model is a giant neural network trained to predict the next word.
- That’s it. One simple objective — applied at extreme scale.
- Scale — trillions of words of training data, hundreds of billions of parameters.
- Out of that emerges reasoning, translation, code generation, summarization, conversation.
- It is exactly Lesson 12’s “model” idea — with a particular shape (transformer) and a particular signal (next-word prediction).
Lesson 13 · Slide 4
The big players (as of 2026)
| Lab | Family | Known for |
|---|
| Anthropic | Claude (Sonnet, Opus, Haiku) | Frontier reasoning, safety focus, long context |
| OpenAI | GPT family | Broadest consumer reach (ChatGPT) |
| Google DeepMind | Gemini family | Multi-modal, integrated with Search & Workspace |
| Meta | Llama family | Open-weights leader |
| Others | DeepSeek, Mistral, Alibaba Qwen, xAI Grok | Frontier or open-weights challengers |
Names and rankings move quarterly — as of 2026.
Lesson 13 · Slide 5
Closed vs open weights (as of 2026)
Closed-weight
Claude, GPT-4-class, Gemini Pro.
- Best frontier capability.
- Accessed via the vendor’s API.
- You do not run the model yourself.
Open-weight
Llama, Qwen, Mistral, DeepSeek.
- You can download and run yourself.
- Great for privacy, cost, customization.
- Capability gap to frontier is much narrower than in 2023.
Lesson 13 · Slide 6
What a prompt looks like
“You are a helpful assistant. Summarize the following email in two sentences and suggest one reply.”
- A prompt is just text.
- Prompt engineering — the craft of writing prompts that get good answers reliably.
- System prompt (instructions, persona) vs user prompt (the actual question).
- The model sees them concatenated, as one stream of text.
Lesson 13 · Slide 7
The context window
The context window is the model’s RAM.
- Everything in one request — system prompt + history + documents + the new question — must fit.
- As of 2026, common windows are 200K–2M tokens (~150–1500 pages).
- Anything beyond the window is invisible — the model “forgets” it.
- Pricing happens here too — vendors charge per million input and output tokens (ballpark single dollars per million as of 2026; verify before quoting).
Lesson 13 · Slide 8
RAG — Retrieval-Augmented Generation
LLMs only know what they were trained on — their training has a cutoff date.
To answer questions about your private documents (or anything newer):
- Retrieve — pull the most relevant chunks from a database.
- Augment — paste them into the prompt.
- Generate — let the LLM answer using both.
The retrieval step typically uses a vector database (Lesson 07).
Most enterprise “ask our docs” chatbots are RAG under the hood.
Lesson 13 · Slide 9
Tools and function calling
A modern LLM can be given a list of tools — functions it may request:
search_web(query)send_email(to, subject, body)query_database(sql)
The model decides one would help, asks the host program to call it, the result comes back into the conversation, and the model continues.
This is how LLMs reach beyond their training data and take action in the world.
Lesson 13 · Slide 10
Agents — LLMs in a loop
An agent is an LLM that calls tools, looks at the result, and decides what to do next — in a loop.
Not magic — a loop. Examples (as of 2026):
- Cursor, Windsurf — IDE agents that edit your code.
- Claude Code — terminal agent that runs commands and reviews their output.
- Devin — brands itself as an autonomous software engineer.
- Operator — web-browsing agent.
- Salesforce AgentForce, Microsoft Copilot Studio — enterprise platforms.
Each one lets the LLM iterate until the task is done — or it gives up.
Lesson 13 · Slide 11
MCP — the agent’s USB-C (as of 2026)
Every agent needs to plug into your tools — Slack, Gmail, the filesystem, a database, the internal wiki.
MCP (Model Context Protocol) — introduced by Anthropic in late 2024 — is the open standard for how agents talk to those tools.
Clients that speak MCP (as of 2026):
- Claude Desktop, Claude Code, Cursor, Windsurf, Cline.
- VS Code (via extension) and a growing list of others.
Write one tool as an MCP server — plug it into any compatible client. This platform has a whole MCP Servers course.
Lesson 13 · Slide 12
The new product landscape (as of 2026)
| Category | Names you’ll hear |
|---|
| Chat assistants | Claude, ChatGPT, Gemini, Copilot, Grok |
| Coding | Cursor, Claude Code, GitHub Copilot, Windsurf, Cody |
| Search | Perplexity, Google AI Overviews, Brave Leo |
| Browsing & action | Operator, Comet, MS Copilot |
| Voice | Advanced voice models from each major lab |
| Image / video | Midjourney, DALL·E 3+, Stable Diffusion, Sora, Veo |
| Open-source orchestration | LangChain, LlamaIndex, LiteLLM |
Every name on this slide is on refresh-checklist.md.
Lesson 13 · Slide 13
Recap
- An LLM is just a giant next-word predictor — at extreme scale.
- RAG and tools let it touch your data and act in the world.
- An agent is that loop, with permissions.
Coming up: Lesson 14 — who does what in IT, the open-source ecosystem, and which course on this platform to take next based on what you’ve enjoyed.