The 2024 AI cheat sheet didn't age well.

Someone sent me an old infographic this morning — "20 AI Concepts Explained," one of those clean little grids that did the rounds in 2024. I started annotating it in the margin. There was a lot to annotate.

01. First reaction

It's a good piece of work. I want to say that up front, because that's what made it uncomfortable to read. Bad outdated material is easy to dismiss. Good outdated material isn't. Someone clearly spent real time on the typography, the layout, the wording of each definition. The squares are well-chosen. The visual logic holds up.

What doesn't hold up is the picture in the back of the reader's head — the implicit map of what AI is that the diagram assumes. That map was accurate in 2024. By May 2026 it isn't.

I'll tell you the moment I knew. I was looking at square 11.

Margin note — Two years ago I had roughly the same picture in my head while I was building the first version of my product. The center of my mental model was the LLM. Today the center isn't the LLM. It isn't even a single point.

02. Square eleven

Square 11 is "AI Agents — autonomous systems that use tools to complete tasks." One square out of twenty.

In 2024 that placement made sense. Agents were ~~a promising application~~ an emerging category — one of many things you could build on top of an LLM, alongside chatbots, summarizers, code completion. They earned a square. Maybe a slightly bigger square than the others, but still a square.

In 2026 that's wrong in a structural way. Agents aren't a category anymore. They're the substrate. When someone says "I'm building an AI system" today, they almost always mean an agent system. Single-shot LLM applications — the kind where you call an API once and return the answer — have become a small minority of what gets built. Not because they were deprecated. Because they were absorbed.

You can date a piece of writing about AI by where it puts agents on the page.

So I picked up a pen and went through the whole grid, square by square. Three labels: still valid, shifted in meaning, no longer earns a slot. I sat with it for an hour. The tally came out to 6 / 12 / 2. Six concepts that survived intact. Twelve that moved. Two that are basically gone.

Six and twelve and two. The number that bothered me was the twelve. You could read it optimistically — "most of it still applies, just slightly different." I think that's the wrong read. The six that survived are foundational concepts (Neural Network, Tokenization, Quantization). They have to survive. You can't shift them; they're load-bearing. The real action is on the floor above. And the whole floor moved.

03. What actually moved

Once I had the twelve laid out, the pattern jumped. Every one of them moved in the same direction. From "how a single model behaves" to "how multiple things work together."

1. Prompt Engineering. Was a differentiator in 2024. In 2026 it's a subset of agent design. You rarely meet anyone whose entire job is writing better prompts; you meet people who design the system that prompts itself.
2. Chain of Thought. Was a clever prompting trick in 2024. In 2026 it lives inside the model. OpenAI's o-series, Claude's extended thinking, Gemini's deep think mode — different names, same admission. Reasoning isn't a prompting pattern. It's a model capability.
3. RAG. Was retrieve-once-then-answer in 2024. In 2026 it's a control loop: retrieve, reason, retrieve again, stop when done. The new label is agentic RAG. It costs more tokens and adds latency, but it solves multi-hop questions that classical RAG quietly failed on.
4. Fine-tuning. Was the first move in 2024. In 2026 it's closer to the last move. You try prompting, tools, and retrieval first. Fine-tune only when those genuinely fail — which is much rarer than people predicted.
5. Temperature. Honestly? I haven't touched the temperature knob in months. With reasoning models, the dial barely matters. The model decides how confident it is.

Then there are the things that need to be on the diagram and aren't. Or are mentioned in passing as if they were footnotes when they should be load-bearing pillars.

The shift, in one line — The 2024 grid was a diagram of the model. The 2026 grid has to be a diagram of the system.

04. What needs to be on the new diagram

Three things. There are more than three, but if I could only add three to the grid, these would be them.

One: protocols. MCP and A2A. I'd argue this is the most important thing that happened in 2025, more important than any single model release. Not because the models got smarter — they did — but because models finally agreed on how to talk. In December 2025, both MCP and A2A were transferred to the Linux Foundation's Agentic AI Foundation, jointly governed by Anthropic, OpenAI, Google, Microsoft, AWS, and Block. By February 2026, MCP had crossed 97 million monthly SDK downloads. That isn't one company's experiment anymore. It's plumbing.

Two: orchestration. Single-agent systems had a short shelf life. The 2026 default is multiple specialized agents, each with a role, coordinated by an orchestrator. Gartner reported a 1,445% increase in multi-agent system inquiries between Q1 2024 and Q2 2025. Fourteen-hundred percent. That's not adoption, that's a phase change.

Three: evals and observability. The unsexy one. Once your agents start acting autonomously, you absolutely have to be able to answer "what did this agent do, why, and on what evidence." If you can't, you can't run it in production. Industry reports put 60% of enterprise AI agent budgets on integration and governance. That's not overhead. That's the product, in a sense. The thing customers buy isn't the model — they assume the model. They buy the assurance that it's auditable.

Margin note — All three of these are baked into how I'm building ATOZAI. The agent-to-tools layer runs on MCP. The agent-to-agent layer runs on A2A. The judgment layer — me — sits on top of evals and observability. That's the architecture. More on that in a separate post.

05. The conclusion that isn't a conclusion

If you read this as a critique of one infographic, you're reading it too small. The infographic is fine. The infographic is a symptom. The actual story is this: the set of things a builder needs to know changed underneath us in twenty-four months, and most of the explainer content hasn't caught up.

In 2024, "knowing how to use the model" was the differentiator. In 2026, it's "knowing how to assemble the system." The models are commoditizing fast. The protocols are standardized. The tools are public. So what's left?

What's left is design. Who does what, in what order, judged against what criteria. And inside that — where the human stays. Not in-the-loop, approving every step. On-the-loop, supervising from above. The shift sounds small but it splits operating models in half.

That's the picture I'm building toward at ATOZAI. This note is the upper layer of that work — the conceptual cleanup before the assembly. The how-it-actually-fits-together part comes next.

— Mr. Ban (@mrban_ai)

---

What I read while writing this

Anthropic — 2026 Agentic Coding Trends Report · VentureBeat — The RAG era is ending for agentic AI (May 5, 2026) · arXiv 2505.02279 — A Survey of Agent Interoperability Protocols (MCP, ACP, A2A, ANP) · MachineLearningMastery — 7 Agentic AI Trends to Watch in 2026 · NVIDIA Developer Blog — Traditional RAG vs Agentic RAG · Deloitte — Unlocking exponential value with AI agent orchestration · MarsDevs — Agentic RAG: The 2026 Production Guide