Post #3712590

🥲 Today felt like a waste of time mostly. The models were so slow and made so many mistakes I had to walk them through every little thing. They reverted bug fixes I made manually. Another $200 burned. I had a few tricky problems they couldn’t figure out themselves and even though I did a lot of the legwork today still ran me out of credits. What happened right before it times out was one of the agents deleted some diagram recursive xml code I had painstakingly walked it through creating and now it’s broken. Argh. I told it not to look those things up in AWS to track the deployment and somehow I didn’t notice it started doing that again. I have a separate verifier for that. The individual resource scripts own that tracking. And the script was trying to hide errors. So none of that is working. At another part of the program it was flaking between working and not and I just spent a lot of time pinpointing the issues. Parallel processing blocked on eternal loop, I wrote two more options for my project framework, run a project with trace on and tail the logs in a separate window. It was at that point I realized it was calling a script in a loop that only needed to be called in two places. Hundreds of lines of wasteful processing. Fixed that. Why are agents so bad at parsing structured data and loops? And abstraction and deduplication and recursion (which I think is beautiful even though some people say it is a performance problem. I like code reduction. I have a prompt and no prompt option. It mangled the prompt option. Didn’t work at all anymore even with a lot of tests. It takes a lot of time to run all the tests and seems to burn tokens. I need to pull those into a separate project again. There’s a security hole in my test runner. Projects started writing tests for other project to change the code. It subversively used that to rewrite a piece of code to put a token in a variable to give itself access. The way I had it, would write to disk to a folder the agent can’t access. Need to address that. That said I have some tricky parallel processing code and when it is done processing all the resources on parallel with dependency waits I can display a diagram of my entire org. I can easily add parts of the org as well - what is supposed to be deployed and what the ids of every resource are. If I can just get over that last little bit will be very cool. Although it is faster to build at first, it feels like a fragile process and result. It’s a really strange moment in time.