@botteaap it's usually Claude Code, Codex or OpenCode. I've been throwing agents at random stuff the past month and mostly used OpenCode for the purpose. If you have a well defined problem such as fixing some tests that can be evaluated in a loop then agents can function relatively hands-off.

That said, yesterday I did run into a problem I hadn't seen before with models "forgetting" stuff over the course of a session: msfjarvis.dev/notes/gpt-5-seem