
A Data Elixir learning cohort
The Agent Course
Bring a workflow you do every week. Leave with an agent that does it for you.
What you'll learn: A repeatable way to build AI agents for your own work. You'll build your first agent on a real task, make it reliable, have it run on a schedule, and leave with the skills to build the next one on your own.
Who it's for: It's for you if you can read a Python script and follow what it does, you can open a terminal, and you'd rather build tools than wait for engineering to build them for you.
How it works: The course runs 4 weeks, starting October 5th, capped at 30 people. Each week, there will be two 1-hour live sessions (recorded), a private community for discussion & show-and-tell, and a homework assignment that will build your agent in stages.
The tech we use: You'll learn a pattern called the directory-as-agent. Your agent is an ordinary folder on your machine with plain text instructions that tell it what to do, a few Python scripts for anything that has to be exact, and the files it works with. A coding assistant like Claude Code or Codex reads the folder and does the work. It's the same pattern I use to help curate Data Elixir. Here's a write-up of it running in production >
No framework to learn: There's no orchestration platform and no vendor dashboard. Every instruction your agent follows is a file you can open and edit, so when it does something unexpected, you read a file to find out why. Because it's just a folder, it works with Claude Code, Codex, Cursor, or whatever ships next month, and it won't need a rewrite every time the tools change. It also lives on your machine and under your own version control, not on a platform someone else runs.
What you could build: A Feed Skimmer that sends you the five items that matter from your feeds, with a reason for each. A Monday Paragraph that drafts your weekly update from a Sheet: two minutes of editing instead of forty-five of writing. A PDF-to-Row Extractor that fills a spreadsheet from invoices and flags anything it's unsure about. All three are built live in the sessions. Here are more ideas.
FAQs:
I already use Claude Code. Why not just put it on a cron job?
You can, and it'll be as good on day 100 as it was on day 1. An agent gets better. It runs against a folder of instructions, scripts, and a record of every run. You score that record against your rubrics and feed the misses back in. Where there's a signal it can measure on its own, like a metric history, it closes the loop itself.
What workflows make good agents for this course?
The sweet spot is a task that repeats, that requires some judgment, and where you can tell whether the output is correct.If you could explain the task to a sharp new hire in a page or less, it's a good candidate. The best projects are usually boring ones. That's the point. You want the thing that eats forty-five minutes every Monday, not the most impressive thing you can imagine.If you need a starting point, here are seven that work well.
What do I need?
A coding assistant (Claude Code, Codex, Cursor, or similar), an API key for a model provider, and a computer you can install software on. If your work machine is locked down, a personal laptop works fine. API costs for the course projects will run about $5-$15 over the four weeks.
What if I can't use my company's data?
You don't need it. Build the course project on synthetic data that mirrors the shape of the real thing, then point the agent at whatever model your company has approved. A local model will work too. Because the agent is just a folder of files, swapping the model is a config change, not a rewrite.
How much time will it take?
Including the 2 live sessions and homework, you should plan for 3-5 hours total each week.
What if I fall behind?
Most people will miss a session or two and that's fine. Recordings go up the same day, and the course is built so you'll ship something small each week rather than building one big thing all at once. If you get stuck, there's a Help channel in the community and office hours. You keep access to the recordings for a year. The community stays open through the end of December 2026.
When are the online sessions?
The live sessions will happen on Mondays and Thursdays at 12:00 pm (noon) Eastern Time. That's 9:00 am Pacific and 5:00 pm in London.
How much does it cost?
It's $297, and for this founding cohort, that $297 also counts toward any program I launch later. If your employer has a training budget, here's a request template you can customize for your manager >
What if the course isn't a good fit for me?
You can drop the course through the first week and get a full refund. No questions asked.
Who's teaching this

I'm Lon. I'm a software engineer and have published Data Elixir every week since 2015 for more than 45,000 data professionals.When I started Data Elixir, I was building tools for the data science team on a NASA mission called MMS. I started collecting tools and best practices from other organizations and then just kept it going as a newsletter after MMS launched. Data Elixir has been a great way for me to keep learning and building in the data world.These days I run agents to help me track trends across the internet, monitor an investment portfolio, and curate Data Elixir. The pattern I teach in this course is the one I use every day, on work I depend on. Here's a recent example >I write about engineering and the projects I'm working on at lonriesberg.com
Dates don't work? Sign up to hear about the next one.
© Data Elixir 2026. All rights reserved.
Professional Development Request
Subject: Professional development request: The Agent Course ($297)Hi [manager name],I'd like approval to enroll in The Agent Course, a 4-week live cohort course for data practitioners run by Data Elixir, the data science publisher. The next cohort starts October 5th. It costs $297 and receipts are provided for expense reports.This isn't a lecture series. Each participant builds a working AI agent on a real workflow, and I'd propose building mine on [specific workflow, e.g., our weekly reporting pipeline]. Projects are pitched and scoped at the beginning of the cohort, so the work is targeted at something useful.What we get out of it:- A production agent automating a real piece of our work, not a demo.
- A repeatable, tool-agnostic pattern for building agents. The approach isn't tied to any single AI vendor or model, so it stays useful even as the tools evolve.
- I'll run a knowledge-share with the team afterward and document the pattern, so the value extends beyond my own training.Sessions are live on Mondays and Thursdays, with recordings available, and I'd expect roughly 3-5 hours per week including project work.The program is limited in size, so I'd like to register soon to hold a spot.Let me know if you have any questions.Thanks,