First five takeaways from AI Engineer Code Summit

I have been thinking a lot about why certain environments make me feel so *alive*. There is this specific feeling I get when I am around people who are operating at a level I aspire to. Every conversation makes me want to grab a notebook, I walk away buzzing with ideas I need to research immediately, and I feel simultaneously humbled and energized. That is exactly what I got from AI Engineer Code Summit last week. ## 🎯 My first takeaway I need to work *extremely* harder at being around people who are much smarter than me. It makes me so happy. I get such an exciting thrill from it. I love being in a room full of people with great ideologies and ideas that teach me new things, where I can take notes mid conversation, go home, and deep dive into everything I just learned. That feeling is so invigorating. ## 🧠 Skills > Agents Barry Zhang and Mahesh Murag from Anthropic completely reframed how I think about building with AI. Stop trying to ship giant monolithic agents. Build reusable skills, which are literally just folders with instructions and scripts. With progressive disclosure, your agent can access thousands of capabilities without overwhelming context. Code becomes the universal interface. ## 🎯 The "Dumb Zone" is real Dex Horthy's Research → Plan → Implement workflow changed everything for me. LLMs get significantly worse around forty percent context capacity. The fix is intentional compaction between steps. Stop polluting context with messy trial and error. Start fresh sessions on purpose. This alone can multiply how effective your agents are. ## ✅ Verification > Specification Eno Reyes nailed it for me. Stop telling agents exactly how to solve problems and instead tell them what correct looks like. The bottleneck is not the models. It is that our code bases are not agent ready. Humans tolerate flaky builds and low test coverage. Agents do not. ## 🚀 The 90 percent vs 100 percent cliff Dan Shipper from Every dropped something that has been stuck in my head. There is a huge difference between ninety percent and one hundred percent AI adoption. Even a small part of your company using traditional methods drags everyone back into that world. His fifteen person team runs four production products with more than seven thousand subscribers, and ninety nine percent of their code is AI written. The secret is Compounding Engineering, where you codify everything into prompts so each feature makes the next one easier to build instead of harder. These are just my first quick thoughts. I will probably write a deeper dive into my favorite talks later.