agent comparison

in the beginning of 2025 i was hand-writing code in cursor with tab complete, occasionally using the agent for more "grunt-work" tasks. i tried claude code when it first came out and didn't enjoy it much, the harness didn't feel as well-developed as cursor which makes sense because it was a lot newer. after that i went over to codex and went all in, buying the $200/month plan and using parallel agents thinking i was being super productive. this was around gpt 5.1 and it turned out to be a bit of a disaster. it wrote a lot of slop and i ended up having to rewrite everything anyway.

when opus 4.5 came out, i gave claude code another shot and version 2.0 was actually good. around the same time i found opencode with oh my opencode and it became my go to for new projects or anything that doesn't have a foundation yet. omo has a bit too much overhead and features i don't really use, but it's great for those less atomic tasks where you need to build something from nothing.

nowadays i'm using codex for my more complex codebases. it catches bugs that opus usually misses and it feels more reliable when there's already a lot of existing context to navigate. the only annoying part is that it asks for permission for every single thing it does, so running it with --yolo is pretty much mandatory if you want to get anything done.

i don't really write code myself anymore. agents still can't build a strong foundation the way i want things structured, but once that foundation is there they can handle a majority of the implementation. i find the only place where i have to go in and really do it myself is frontend related as these models don't really have good taste yet. i've learned a lot about this way of working from steipete's blog post about his setup and his agent-scripts repo.