La Sagrada Familia, GTA 6, that bun post about the rust rewrite.

For my Spanish audience. That's all I have to say about the too-dangerous-to-release newest Anthropic model.

Loops that prompt agents to build programs? Nah, One builds Hoops that build Loops that prompt agents to buid programs. And there is no reason to stop there. Coils can build Hoops, Frolicks can build Coils, and so on. If you are not building a fractal of factories that build factories that prompt agents to write code, you are falling behind.

I read people too much. In a casual conversation, with most people, I immediately notice why they are saying what they are saying. The tone, the non-verbal signs, the fakeness (specially compliments, which I despise), etc. I get impatient quickly, because I feel like I’ve already decoded the intention, so I cut people off a lot. Which makes me look awkward, or even rude. People think this is because I’m an engineer, and that automatically puts me in the “bad at social interactions” box. But I don’t think that’s it. I think I might just be oversensitive. That being said, this awareness does not translate to how I speak myself. I’m extremely easy to read. It’s very hard for me to lie, and I often fumble, fastly chaining one word after another with no plan whatsoever, which sometimes takes me to weird places, like an LLM. Does this feel familiar?

Yesterday I woke up in the middle of the night, sweating, heart racing. I grabbed my phone and immediately bought anthropig.com. Now I have two questions: - What should I build with it? - And will I go to jail if I do?

Three personal (and probably biased) predictions: - SaaS will keep growing. Growth hasn't slowed but accelerated on the last 3 years. Turns out companies don't want to build nor host things that are not critical to their businesses. Not even if vibe-coding lowers that cost. Another example of misspriced "everyone will not just." in the markets. - Companies will interact with SaaS through a single external cloud agent. Every SaaS is building their own agent, but people want to do tasks that work across many different tools. Also, computer use is a hack, an IT workaround. It will be replaced by cloud agents as soon as they are ready (the challenge being auth and scraping). - Jobs are not going to disappear, the world is not going to end, and things are going to change less than we think.

I wrote a new article, exploring the consequences of pain and pleasure on LLMs. I had a lot of fun writing it, so don't take it too seriously. paoramen.fika.bar/agentic-pl…

Why is everyone talking about 100 eggs developer

I can't believe it... Handinger can play wordle? yeah, and is pretty good at it!

Don't let the thought leaders fool you. Harness Engineering is just Helicopter Parenting for LLMs.

If you believed newsletter media kits, rich people and C-level executives do nothing but read newsletters.

In a world where CEOs are pushing token maxxing, I decided to instaurate token fasting. At least once a week, we code by hand. Just fingers and brain madly dancing together once again. And it rhymes. Not only that, quarantined one of the most delicate organs in the whole beast: Handinger’s harness. Why? 1. Because humans are tactile animals. This is why we take notes in school and learn to drive in an actual car. I need us to understand the codebase in our bones. The dependencies, the seams, the bugs and dreams. And it rhymes again. Reviewing generated code gives you the dangerous feeling of understanding. But it is mostly shallow. You nod along, accept the diff, and walk away convinced you know what happened. 2. Humans already have a disease: when solving problems, we add things instead of removing them (there is a famous study about it, go read it!). Clankers do this 100x worse. They feel no pain, so they add forever. Another file. Another method. Another “helper” that, honestly, could have been inlined. A growing pile of computational debris. In some areas of the product, I don’t mind. CRUDs, APIs, UI. Fine. But a harness is a zero-sum domain. Every new tool, prompt, and abstraction has a cost. It decreases margins and effecteviness while increasing latency. This is why, paradoxically, Claude Code has now turned into a mediocre harness for their own models. I'm not being anti-AI, but everything interesting in the universe (like life itself) tends to happen in between.

I think I know how to solve the lack-of-understanding problem when using slop machines. Add a pre-commit hook where an LLM asks you 3 questions about the code you are about to commit. If you can’t answer them correctly, it doesn’t let you commit.

Today I helped a friend automate part of his company. He has a freediving center and wants to sync the booking software with his mailing list. Basically: read a booking email, extract the information, and create the contact to the mailing list provider. I suspect every company has several workflows like this. People copying and pasting information between systems that don’t speak to each other. Since this is a deterministic workflow, I thought of trying n8n and make. What a drama. Regexes, JSON nodes, variables, and truly piles of complexity for a relatively simple task. I struggled for 3 hours, and after watching several videos about "how to make a converger node" I gave up. How are non-coders supposed to use this? Then I thought: “Maybe Handinger can do this?” We haven't designed it to do this kind of workflows, but YOLO. We created a worker, integrated it with his mailing provider, and forwarded one of the emails to the worker’s inbox. It...worked!? LOL, I can’t believe it. I'm used to things not working as expected, not the other way around. This is, in 20 years, the first time I’ve built a product that can do more than I thought it could.

@threepointone built with cloudflare agents!