What I and many seem to want is a way to truly put LLMs to work, ways to outsource higher volumes of work without explicit instruction/live oversight.

I think this is a fascinating problem, and in many ways how I hope to use gptme, but not really there yet.

Ideas:

• ask it to try implementing feature/fix/refactor x, then typechecking/testing/linting it, then make a PR if successful
  • if unsuccessful, attempt up to n retry strategies, possibly in a branching manner where we try to detect the best branch, since any step could go wrong and put it in an unusable state.
    • started working on this in refactor: work on programmatic interface, self-reviewing agent #199
  • Similar goals to SWE-Bench stuff: feat: started working on SWE-bench evals #142
• various automation workflows, putting it in a cron job, etc.
• return to the gptme roots and give it your documents/data/activitywatch to inspect and review, condensing it into a daily report or similar
• browsing the web for content, giving the user recommendations
• write docs for how to do all these things, give the docs to gptme so you can ask it to write scripts doing it
  • Minimal docs added to https://gptme.org/docs/automation.html in docs: added minimal automation docs #144
• more subagent examples/evals
• feedback loops, like make edit to page -> view web page -> make edit
  • Used it for generating https://github.com/ErikBjare/gptme/blob/master/site/style.css with much success

What we should focus on:

• Making gptme work on gptme while I sleep

What we'd need:

• Tool customization
  • It is likely that performance would improve by limiting each step to a restricted set of tools (would at least save on tokens)
    • Implemented --tools option in 48d559b
• Reliable message passing between gptme runs/agents
  • The current subagent mechanism of asking it to output JSON isn't the most reliable.
    • Maybe making it a prompt chain would fix it? (done in 0dd6583)

Issues:

• How to avoid burning lots of non-value generating tokens while I sleep?
  • Make reliable, do intelligent tree-search of branches, limit retries