The Changelog: Software Development, Open Source - Fresh beats for endless flow state (News)

Episode Date: April 21, 2025

We drop our fourth Changelog Beats album, Dex Horthy proposes the 12-factor AI agent, Thorsten Ball takes us step-by-step through building a coding agent, Zachary Huang builds an LLM framework in 100 lines of code & Philip Laine's Spegel project gets unknowingly forked by Microsoft.

What's up nerds? I'm Jared and this is changelog news for the week of Monday, April 21st, 2025. A recently published study called The Effect of Deactivating Facebook and Instagram on Users' Emotional State proved what we all knew to be true, but previously didn't have a study to point to. Social media, at least in the form we've been using it for the past decade plus, is a total drag. Oh well, let's get into this week's news. Fresh Beats for endless flow state.

Our fourth full length change Log Beats album dropped today. It's called After Party and it features beloved Breakmaster Cylinder tracks from our outros, ad rolls and transitions. It's essentially 26 chill beats to help you get into a state of flow and stay there. Here's a sample of a few of my favorites, such as Track 5, Severanced Package.

Track 11, do not stare directly into the changelog. Track 25, Sunk and Barge Zone, Adam's sudden Death version. And track 26, Last Light. light. If that one's familiar to you, yep, it's our long time outro to change log interviews which many, many people have asked us to make available to stream or buy. Well, now you can. Head to changelog.com slash beats for all the links to all the music platforms. Or simply search for changelog beats in your favorite music app and we'll see you at the after party. The 12 Factor Agent.

Dex Horthy, who has been hacking on AI agents for a while, set out to answer the question, what are the principles we can use to build LLM powered software that is actually good enough to put in the hands of production customers? What he came up with is 12 factors in the spirit of the 12 Factor app. Here they are. 1. Natural language to tool calls. Two, own your prompts.

Three, own your context window. Four, tools are just structured outputs. Five, unify execution state and business state. Six, launch, pause, resume with simple APIs. Seven, contact humans with tool calls. Eight, own your control flow. Nine, compact errors into context window. 10, small, focused agents.

11, trigger from anywhere, meet users where they are. And 12, make your agent a stateless reducer. If those 12 bullet points are more confusing than enlightening, that's because, hey, they're just bullet points. Click through for the full explainers. How to build an agent. Speaking of agents, you might now be thinking

it's super hard to build one. Thorsten Ball says you're wrong. Quote, it's not that hard to build a fully functioning code editing agent. It seems like it would be when you look at an agent editing files, running commands, wriggling itself out of errors, retrying different strategies. It seems like there has to be a secret behind it. There isn't. It's an LLM, a loop, and enough tokens." Turns out you can build a small

and yet highly impressive agent in less than 400 lines of code. Thorsten proves this out by building an anthropic-based agent in Go over the course of this blog post. I followed along and I have to say, it's all very basic stuff. Thorsten's closer? Quote, these models are incredibly powerful now. 300 lines of code and three tools and now you're able to talk to an alien intelligence

that edits your code. If you think, well, but we didn't really. Go and try it. Go and see how far you can get with this. I bet it's a lot farther than you think. And that's why we think everything's changing. It's now time for sponsored news. Teams with faster builds ship faster.

Depot CEO, Kyle Galbraith, recently shared his thoughts on the age-old question, build or buy for CI-CD. The fact is teams with faster builds ship faster and more, and teams that know how to CI well create faster builds. Imagine what would happen if your builds got 10x faster. Kyle found that when companies talk about building their own CI, what they actually mean is self-hosting runners from another CI provider.

After analyzing hundreds of engineering teams, he identified three distinct archetypes. First, there's the abstraction first team. These folks know what matters and move fast. They discover Depot and immediately think, it's 10 times faster, half the price of GitHub actions and I just changed one line of code? No-brainer. I discover Depot and immediately think, it's 10 times faster, half the price of GitHub Actions, and I just changed one line of code? No brainer. Then there's the Infracurious team.

They're in that honeymoon phase with self-hosting. GitHub Actions is slow, so they spin up their own runners in AWS. But now they own the uptime engineering, literally, and they now have to learn how to make the builds faster and keep them secure. These folks eventually become the third and last archetype which I'll let Kyle cover in his post.

It's linked up in your chapter data and the companion newsletter. Thank you to Depot for sponsoring changelog news. Check them out at depo.dev. Pocket Flow is a 100 line LLM framework. The author of Pocket Flow, Zachary Huang, thinks current LLM frameworks like Langchain, Crew AI, Langraph, etc. are all bloated. After reading Thorsten Ball's essay on how to build an agent, I can believe it.

The 100 lines in Pocket Flow capture the quote core abstraction of LLM frameworks and you build on top of that to do multi-agent, workflows, rag, et cetera. One great example of Pocket Flow in action is this repo that's linked up in your newsletter, which crawls GitHub repositories and builds a knowledge base from the code. It analyzes entire code bases to identify core abstractions and how they interact,

and transforms complex code into beginner-friendly tutorials with clear visualizations. What a great way to learn a new code base. Getting forked by Microsoft. As a sole maintainer of an open source project, Philip Lane was excited when Microsoft showed interest in Spiegel, his tool to enhance Kubernetes cluster scalability through peer-to-peer image distribution.

But then things got weird. Weird, weird. At KubeCon Paris, Lane attended a talk about strategies to speed up image distribution that mentioned a Microsoft project called Peer D. Quote, while looking into Peer D, my enthusiasm for understanding different approaches in this problem space quickly diminished.

I saw function signatures and comments that looked very similar as if I had written them myself. Digging deeper, I found test cases referencing Spiegel and my previous employer, test cases that had been taken directly from my project, references that are still present to this day. The project is a forked version of Spiegel maintained by Microsoft, but under Microsoft's MIT license." Spiegel is also MIT licensed, which does allow for forking and modification without contributing back,

but does not allow, quote, removing the original license and purporting that the code was created by someone else. End quote. Not cool, Microsoft. What's an open source maintainer to do in this circumstance? Philip says, as an effort to fund the work on Spiegel,

I have enabled GitHub sponsors. This experience has also made me to fund the work on Spiegel I have enabled GitHub sponsors. This experience has also made me consider changing the license of Spiegel as it seems to be the only stone I can throw. That is the news for now but also scan the companion changelog newsletter for even more links worth clicking on such as, Xmorelda is an AI chatbot just for Elixir. Pipelining might be my favorite language feature. And an awesome collection of GPT-40 images and prompts. Get in on the newsletter at changelog.com slash news.

Do it, do it. We have some great episodes coming up this week. On Wednesday, Kendall Miller from CTO Lunches joins me. And on Friday, Matt Ryer takes Adam and I to Matt World. Have a great week, like, subscribe, and leave us a five star review if you dig the show. me.