AI-DEV 100: Build an Agent Before You Have an Opinion About Agents

Most people's mental model of AI agents comes from watching one work. A demo runs, the output appears, and the mechanism stays invisible. That model holds up in conversation and collapses the moment you have to decide whether AI can take a task off your own plate, or explain to a colleague why the thing they saw in a keynote will not do what they think it will do.

AI-DEV 100: Agentic Workflow Workshop closes that gap the only way it reliably closes. You build one.

Four modules, no prerequisites, roughly an hour of lecture and demo each plus an hour of hands-on labs.

Who this is for

You work in a non-engineering role where the week fills up with repetitive knowledge work: research, writing, formatting, and copying things between tools. You already know how to chat with an AI. You have never built one that runs on its own. You are curious rather than anxious about it, but you want to know what these systems actually do, not what a keynote says they do..

You do not need to write Python. You do not need to pick a framework. Everything you build in this workshop is written in plain markdown, saved in a folder on your own machine, and run in Claude Code, Cursor, Claude Desktop, or ChatGPT.

The modules

1. The Mindset Shift

The vocabulary module, and the one that makes the rest coherent. The difference between assistive AI and agentic AI. Four patterns on one spectrum: Chat, Collaborate, Workflow, Automation, and how to tell which one a given task deserves.

You learn the four components every good one has: Role, Context, Constraints, and Output Format. Then you learn to test it, because a prompt that works once and works differently the next time is not a system.

Lab: write a system prompt that scores records against a set of criteria.

2. The Research Agent

The first principle here is that judgment has to be written down before it can be delegated. A task you do by instinct twenty times becomes an agent only once you can state the criteria, weight them, and say what evidence counts. Vague standards produce vague output at scale.

The second is verification. An agent's failure mode is not a stack trace, it is a plausible-looking wrong answer delivered with total confidence. So you make the work auditable: the agent shows its sources, and you sample a few results and check them against reality. A few minutes of sampling tells you whether to trust the whole batch, which is the only thing that makes running a batch safe.

Lab: hand an agent your criteria and a stack of work. Check its homework.

Lab: build and run a working agent.

3. Chaining Multiple Agents

Why one agent is not always enough, and why adding more is not automatically an improvement. Orchestrator and subagent patterns. Routing, handoffs, and context sharing. Failure modes specific to agent networks, where a small error in one agent's output becomes another agent's confidently wrong input.

Lab: one input in, a week of output back, in your voice.

4. Agent Workflows

Most people build one good agent, feel great about it, and never open it again. This module decides whether anything you built survives the week after the workshop.

An agent pays off only when it hangs off something that already happens. So you start from where your time actually goes, automate what repeats, and give every agent a trigger you cannot forget: a routine, a ritual, or a schedule.

Then one agent stops being enough, and you find out that adding more does not automatically make things better. You split the job across an orchestrator and its sub-agents. You hand one agent's output to the next, and you see why a chain of small agents beats one prompt carrying every job at once. You also watch a small error in one agent's output become the next agent's confidently wrong input.

Lab: Decide what runs without you, and where you stay in the loop.

Format

Four modules, each with roughly an hour of lecture and demo, plus an hour of hands-on labs in total. Reference materials are yours to keep. No prerequisites.

AI-DEV 100 can be scheduled for your team. If your work is closer to the hardware AI runs on than the applications built on top of it, look at AI-INFRA 100: GPU Cluster Infrastructure instead. The two tracks are independent, and neither is a prerequisite for the other.