Every Coffee and Open Source episode comes with the same pile of chores. Stage the guest on the website, write a bio, track down the Spotify, Apple and Amazon links, then publish. Being the lazy developer I am, I handed those chores to an agent, with some pretty strict rules. Don't invent facts. Show me the diff and stop. Only use a link if it's an exact match.

Then it staged the guest with the bio of a completely different person with the same name. In its defense, it didn't make anything up. It found a real person, just the wrong one. I caught it before the episode was published (sorry, other person), and it got me thinking about where a human actually needs to be in the loop.

On my site, one commit publishes an episode and posts it to social media, so when the coding agent publishes one, it opens a pull request and I'm the one who merges it. The .NET dashboard I built to prepare episodes can't post or push anything, and every preview needs its own approval, tied to a hash of what I reviewed. The app I'm building to stream the show won't go live until I type GO LIVE.

In this session we will go through the real skills and approval gates, mistakes included. You'll see the difference between a rule an agent is asked to follow and a gate it can't get around, and why I'd take a tool that fails loudly over one that quietly fakes a step.

Everything in it runs on my own podcast tooling, and the website repo is public at https://github.com/isaacrlevin/CoffeeAndOpenSource.com.