I finally tried Claude Code for most of last week on a toy Typescript project of moderate complexity. It's supposedly the pinnacle of agentic coding assistants,

I finally tried Claude Code for most of last week on a toy Typescript project of moderate complexity. It's supposedly the pinnacle of agentic coding assistants, and I tend to agree, finding it far ahead of Copilot et al. Seeing it working was like a bit of magic, and it was very addictive. It successfully distracted me from my main projects that I code mostly by hand. That said, and it's kind of hard to express this well, not only is the actual productivity still far from what the hype suggests, but I regard agentic coding to be like a bad addictive drug right now. The promise of magic from the agent is always just seems around the corner: just one more prompt to finally fix the rough edges of what it has spat out, just one more helpful hint to put it on the right path/approach, just one more reminder for it to actually apply everything in CLAUDE.md each time... Believe it or not, I spent several days with it, crafting very clear and specific prompts, prodding with all kinds of hints, even supplying it with legacy code that mostly works (although written in CSharp), and at the end it had written a lot of that almost works, except a lot of simple things just wouldn't work, not matter how much time I spent with it. In the end, after a couple of hours of writing the code myself, I had a high a quality type design and basic logic, and a clear path to implementing the all the basic features. So, I don't know, for now even Claude seems mostly useful only as a sporadic helper within small contexts (drafting specific functions, code review of moderate amounts of code, relatively simple refactoring, etc). I believe knowing when AI would help vs slow you down is becoming key. For this tech to improve, maybe a genetic/evolutionary approach would be needed. Given a task, the agent should launch several models to work on the problem, with each model also launching several randomized approaches to working on the problem. Then the agent should evaluate all the responses and pick the "best" one to return.