I was recently speaking to a very dear friend who isn't a dev but has an itch to build, and he's been learning tools quickly, helping him vibe code fun things like a personal site, a fidget spinner that responds to music and such.
The issue is that Dunning-Kruger effect catches up quickly. He now wants to build a "coding agent". The difference is that this one will be 'end to end' - you give it an idea and it will deliver the final product.
Since he stays on top of tech news, he is aware of the Claude Code leak and what he supposedly gathered from it is that Claude Code is just a bunch of .md files. (He straight up asked me "Have you heard of an md file?")
So his understanding is that an agent essentially just needs to be given the right set of instructions in md files, and since no one else has figured how to do that right (until he now will), we don't have agents running end to end.
Technically, you can ask an LLM to do something "for each entry made by the user" instead of using a deterministic for loop. You can write your conditionals as pure English if statements. But the cost in terms of tokens, time and deterministic results is something that you will never obviously understand you are missing out on.
After speaking to him, I felt a sense of ease. Non-technical people will make simple apps but they are not going to build everything themselves. Simpler systems will still 'work', as long as the cost of time, tokens and lack of deterministic results are not too noticable or at least forgivable. Anything bigger will break.
Top comments (0)