Self-enhancing agents

[BradKML]

Got this idea while pondering the scientific feedback paper, maybe agents can spawn derivative agent(s) that has modified scripts such that they can better think about the work that they do? https://github.com/Eladlev/AutoPrompt Weixin-Liang/LLM-scientific-feedback#32

[SmartManoj]

Yes.

[BradKML]

If this is true then, what would be the ideal workflow to pre-tune SWE agents to "get to work" better?

1. Get the LLM to start with self-optimizing with subsets of SWE-Bench Lite, finding the best prompts and cheat sheets
2. Use the self-taught information to punch through other hardcore benchmarks like LiveBench
3. Profit???

Can't seem to figure out what can be missing from this framework, like cheat sheet RAGs maybe?

[SmartManoj]

pseudocode; exit_code discussion

def do_task(task):
    success = False
    while not success:
        success = command(poetry run python -m openhands.core.cli -t "task").exit_code == 0
        if not success:
            do_task('Analyze the error in {log_file} and fix it')


do_task('some task')