Within the next five years, I want to fly into space with @SpaceX as a tourist and spend a few days there. Seeing Earth from up there would be a dream come true. 🚀🌍

„you” can just build things

Whenever I am in a conversation with AI, I often end my sentences with "You know what I mean?" Anyone else? 😁

Today I told someone who knows nothing about AI, "Dont hallucinate like an LLM." He just stared at me like I was from another planet. 😂

I think every software engineer has gone through a phase of imposter syndrome at some point. I had not felt it for a long time, not even since I started with agentic development. But in the last 3 weeks, that feeling has come back again and again. I even feel it right now. I feel like I am building things that would not have been possible before. And this time, I actually enjoy that feeling 🌸

Many thanks for github.com/anthropics/sandbo… @AnthropicAI ❤️ It is a good combo with function-hooks and Claude-Mods

Since November 2025, I have been building software only with agents. Since then, my expectations for good software have grown a lot. But one thing has not changed. It still feels like building good software takes just as much time as before 😁

I think the day the average user of agent tools like Claude or ChatGPT truly understands the difference between stochastic and deterministic, we will start talking less about how to write better prompts and more about how to verify the results.

👀

With tools like REA, I think the best way to protect your core implementation is to stop shipping it to users and move it to the cloud. Of course, people can still reverse-engineer the APIs and behavior, but they can not reverse-engineer the server-side code itself. github.com/morluto/rea

I have been deep in the sandboxing rabbit hole for about 3 weeks now. Playing around with function hooks and Claude Mods (long before it was official). The fact that we can now interfere so deeply with how Claude Code works opens up a whole new world of possibilities. Really exciting stuff! Many thanks to the Claude team, @bcherny @trq212

🔥🔥🔥

Mark my words: this will happen sooner than most people think. At some point, maybe not even that far away, some of the biggest stars in the world will not exist in real life at all. Completely virtual artists, actors and personalities with millions of fans. And people will go crazy for them the same way people once did for Michael Jackson or Tupac, and still do for today’s biggest stars. There will be groupies, fan clubs, scandals, relationships, everything. Just for someone who was never a real person.

For a while now, I have been letting Claude continue working during my lunch breaks. I give him a few tasks, tell him to coordinate himself, work around problems if needed and report back when I am back. So… does this still count as a lunch break? Technically the work never stopped 😄 🙃

Finally 🎉 @typesafeai thank you so much 😊

Next time, I am complaining to Skynet 😁

Here is my write-up on what AI changed about software engineering and what it did not. I revised it in several passes, and it comes down to four things: 1. the work moved one level higher, 2. most of the new agent vocabulary is old engineering, 3. the same tools on a weak foundation make things worse, and 4. juniors still have a path to senior. It is a 15-minute read: honey.glass/blog/code-is-che… The work moved one level higher. We design the systems the code lives in and set the rules AI has to work within. From up there, we look straight down at the software and still see the users, the product, and the business. Most of the new agent vocabulary is old engineering. Graph-Engineering is state machines or workflows. Loop-Engineering is loops, retries and control flow. Harness-Engineering is the execution layer around agents. The context is new. The concepts are not. The same tools on a weak foundation make things worse. Give an experienced developer a solid codebase, and the improvement in quality and productivity can be huge. Put the same tools on a weak foundation and the result is more problems than they solve. Juniors still have a path to senior. If they write less boilerplate, they can learn by reviewing AI-generated code instead. But they should still build things themselves and make mistakes. I wrote all of this down after I skipped the engineering myself and lost several days to it. Generating code has become cheap. Building good software has not. If you see it differently I would like to hear it. That is half the reason I published it.

☹️ @AnthropicAI