South Summit 26 Talk: Beyond the Buzzwords, AI Agents and the Modern AI Stack

An AI agent is not magic. Strip the word back and what’s left is a reasoning brain, a loop, and a handful of tools. Once you can see the loop run, the magic turns into engineering. The trouble is that “agent” arrives wrapped in a pile of other buzzwords, LLM, RAG, MCP, that all blur together. This is the map: what each term actually does, and how they fit into the modern AI stack.

This is the deck from my South Summit 2026 masterclass. Scroll through it below, or open it fullscreen.

An LLM is a next-token predictor, which is why it needs tools

A language model reads text as tokens, turns them into numbers, and runs them through a deep stack of transformer layers to guess the next token. Everything it “knows” is frozen in its weights at training time, and it can only hold so much in mind at once: the context window. The newest models stretch this far, the latest Claude reaches 1,000,000 tokens, roughly 750,000 words, but it’s still finite: past that edge, older text falls away.

Two limits follow directly from this design. The model has no live data, so it can’t tell you today’s news, today’s prices, or anything past its training cut-off. And when it doesn’t know, it doesn’t stop; it autocompletes something plausible. That’s hallucination, and it isn’t a bug you can prompt away. It’s what a next-token predictor does when the honest answer isn’t in its weights.

This single gap, a brilliant reasoner with no senses, is what every other piece on this page exists to close.

RAG grounds the model in your data

The first way to close the gap is to bring the data to the model. Retrieval-augmented generation fetches the documents relevant to a question, pastes them into the prompt, and lets the model answer with them in front of it. It’s how most “chat with your docs” products work.

RAG:  question → retrieve → augment prompt → generate → answer

The model still just predicts the next token, but now it predicts over your data instead of only its frozen weights. Retrieval is a step you bolt in front of generation, so the answer is grounded in sources you control rather than the model’s memory.

MCP is one protocol between models and tools

The second way to close the gap is to give the model tools it can call: search, a database, an API. The problem is that every tool speaks its own dialect, so each model-to-tool connection used to be bespoke. The Model Context Protocol fixes that with a single interface. Build the integration once, and every MCP-capable app can use it.

MCP is to AI tools roughly what USB-C is to devices: a standard plug. It’s what makes an agent’s tools pluggable instead of hand-wired, and it’s why tooling is the layer of the stack moving fastest right now.

The loop is what makes it agentic

Now combine the pieces. Give the reasoning model a set of tools and let it run in a loop: reason about the question, decide whether to call a tool, read the result, and repeat until it can answer. That pattern has a name, ReAct (Yao et al., 2022), and it’s small enough to draw.

flowchart LR
  U([Question]) --> LLM{LLM<br/>reason}
  LLM -->|need data| TOOL[call a tool]
  TOOL --> LLM
  LLM -->|done| A([Answer])

The model owns this loop: it chooses which tool to call, how many times, and when to stop. That handoff of control, from you to the model, is the entire definition of “agentic.” Everything else is plumbing around it.

The modern AI stack is a span, not a box

This is why “AI agent” resists a one-line definition: it isn’t a single component. It spans the model (the brain), memory (the context it carries between turns), tools (its reach into the world), and orchestration (the loop that ties them together). The buzzwords each name one slice of that span.

Read the table top to bottom and the jargon stops competing and starts composing. An agent is what you get when you assemble the whole column and let the model drive.

The takeaway is simple: an AI agent is a brain in a loop with tools, and the buzzwords around it are just names for the parts. Once you can see the loop, the magic becomes engineering.