Lesson 2: Meet the AI Agent

In Lesson 1 we said agents do work. Now let's open one up. The good news: you don't need code to understand an agent. You need an analogy, and Jensen hands you a great one.

The old way vs. the new way

For 40 years, using a computer meant the same ritual: launch an app, click around, type things in. Jensen says that's ending:

"Whereas we used to launch an application, click and type, we now replace that with explaining to the AI what we want — our intent — and the AI generates the code, or uses tools, and produces the necessary output."

— Jensen Huang, NVIDIA GTC Taipei 2026

This is intent-based computing. You describe the outcome you want ("turn these receipts into a summary," "build me a CAD file for this part"), and the agent figures out the how. Notice the echo of our other module: you express the outcome; the technology handles the details.

Anatomy of an agent: brain, body, workshop

Here's the mental model to keep forever. Jensen breaks an agent into parts and gives each a human metaphor:

"You can think of the model as the brain, the harness as the body, and the tools that it uses, working in a runtime — think of it as a workshop."

— Jensen Huang

Unpacking that into the five pieces every agent has:

Model (the brain) — the large language model that does the thinking: understanding, reasoning, planning.
Harness (the body) — the software that orchestrates everything: routes information, manages memory, decides which tool to use when.
Tools — the real things it can use: a web browser, a spreadsheet, a database, a payment system.
Skills — basically instruction manuals that teach the agent how to use each tool well.
Runtime (the workshop) — where it all actually runs (a cloud, a server, your PC).

And like us, an agent juggles two kinds of memory — working memory (what it's focused on right now) and long-term memory (what it remembers across time). It runs a loop: observe → reason → plan → act, over and over, until the job is done.

The one-sentence version: an agent is a thinking model (brain) wrapped in orchestration software (body) that uses tools (in a workshop) to get a goal done — managing memory along the way.

Why this rewrites software itself

Jensen's bigger claim: the agent isn't just a new feature — it's the new application. The old stack was "software running inside an operating system." The new stack is an agent running in a runtime:

"The application… is going to be replaced by an agentic runtime. And that is the modern application: an agent."

— Jensen Huang

That's a genuinely big deal for anyone who works with information systems. The fundamental building block of business software is shifting from "an app you operate" to "an agent you delegate to." And he's emphatic about how widespread it gets:

"Every company will be an agent company. Every company will have agents running inside."

— Jensen Huang

You've studied systems as inputs → process → outputs, with databases and interfaces. The agent is a new architecture on top of all that — and the five-part anatomy is your checklist for evaluating or scoping one. Which model? What tools and data can it touch? Where's the runtime? What's in long-term memory, and is that data sensitive? Each part is also a control point: tools define what an agent is allowed to do, and the harness is where you enforce permissions. Knowing the anatomy is how you move from "AI is magic" to "here's exactly what we're deploying and where the risks are."

Quick check

In Jensen's analogy, what is the "harness"?

Model = brain (thinking), harness = body (orchestration), runtime = workshop (where it runs). The harness is the coordinator that turns a smart model into a system that actually gets work done.

🧠 Think like an MIS analyst

Design an agent for invoice processing

Your finance team spends hours keying invoices into the accounting system. You're asked to sketch an agent for it. Using the five parts, what does it look like — and where does a human stay in the loop?

Model (brain): an LLM that reads invoices and decides what to do. Tools: an email/inbox reader, an OCR tool to extract text, the accounting system's API to create entries, the vendor database to verify suppliers. Skills: manuals for "how to read our invoice format" and "how to enter a bill correctly." Runtime: a secure company server. Memory: long-term memory of past vendors and corrections. Human in the loop: a quality gate — anything over $5,000, or any new/unknown vendor, gets routed to a person for approval. Notice you just scoped a real system without writing a line of code, because you understood the anatomy.

✍️ Your turn

Pick a repetitive task from your life or a job. Name the agent's tools (what real things would it need to touch) and where you'd put a human checkpoint.

"Every company will be an agent company." Name one specific way your school, job, or favorite company might run an agent in the next year.

Key takeaway: An AI agent = model (brain) + harness (body) + tools & skills (in a runtime/workshop), running an observe-reason-plan-act loop with memory. It's becoming the new "application" — and knowing its parts lets you scope, evaluate, and govern real systems instead of guessing.

← Previous1 · Useful AI Has Arrived Next →3 · Compute Is Revenue