Hi everyone,
I recently did give a try to a coding agent, seeing the fuzz about it everywhere finally convinced me to click on it. I briefly tested it for scientific code translation from Fortran but already face some issues:
- It does well on very simple tasks, but it often makes many small, stupid mistakes (such as stopping working in the middle of a function, a in the next step, erase everything…)
- One annoying behaviour: it runs Julia with commands like
julia --project -e "some code here"
and that creates shell issues for me (namely because of the!in some function names). The agent often does not accept that the code is actually correct. - Sometimes it kills the Julia process before the program finishes and then concludes the code is wrong.
- Right now I use the basic GitHub Copilot-(pro/teaching version) agent in VSCode (using Claude Sonnet 4). That combo seemed to handle Julia better for me than GPT-4, but I have not modified any agent settings (I am not really an AI guy, and don’t know all the tricks…)
I know you can give custom instructions to these agents, but I could not find a thread where people share their setups (hope this is not a duplicate).
Could you share:
- Your experience with coding agents, especially in the context of Julia code ofc (good and bad)?
- Your custom instructions, prompts, or agent settings that help with Julia development?
- Any practical tips to avoid the problems above (shell runs, premature killing, false negatives)?
Thanks !