Track 3
Agents and Tooling
How models take action and interoperate.
The Agents and Tooling track covers how AI systems stop talking and start doing. An agent takes actions with side effects, which changes the engineering questions from answer quality to safety, specification, and interoperability.
Four modules map the territory. How AI Agents Work explains the core architecture: the language model as reasoning engine, tools as hands, memory, the ReAct loop, and the guardrails that keep irreversible actions behind checkpoints. How to Design an AI Agent lays out a ten-step evolution from a manual runbook to orchestrated multi-agent systems, with the human accountable at every stage. How MCP Works introduces the Model Context Protocol, the open standard that replaces bespoke integrations with a host, client, and server architecture carrying resources, prompts, and tools. How ChatGPT Apps Work shows the same protocol powering interactive widgets inside a conversational surface, along with its strict sandbox security model.
Read this track before approving or building any system that acts on production data. The recurring theme is that guardrails belong in code, not in prompts.
Modules in this track (4)
How AI Agents Work
Agents are where AI stops talking and starts doing things that change the world.
How to Design an AI Agent
Agents are built through a ten-step iterative evolution, not a single deployment event.
How MCP Works
The Model Context Protocol is the emerging standard for connecting AI systems to data and tools, and it is already in production.
How ChatGPT Apps Work
The distribution surface has shifted. Eight hundred million weekly ChatGPT users can now use interactive third-party apps inside the chat.
How this track fits
The AI Learning Hub is built as a sequence rather than a library. Foundations covers what the models are and how they behave. Context covers the material a model needs around it before its answers can be trusted. Practice covers what changes when a system leaves a demonstration and meets real users, real data, and a real obligation to be right.
Tracks can be read in any order, but the sequence matters more than it looks. Most failures in institutional AI are not model failures. They come from deploying a capable model into an environment that was never prepared for it: no owner, no stopping condition, no measure agreed in advance. The later tracks exist because the earlier ones are not sufficient on their own.
For the foundations underneath all of this, the free Learn LLMs course works from next-token prediction up to the Transformer, and LLM Architecture Explained covers the same ground visually. To move from models to systems that act, see Learn Agentic AI. For the governance that has to sit around any of it, see AI governance in the United Nations and the governance templates.