[jasonsteving]

You nailed an interesting nuance there about agents needing to make their own decisions!

I'm getting fairly excited about "agentic" solutions to the point that I even went out of my way to build "AgentOfCode" (https://github.com/JasonSteving99/agent-of-code) to automate solving Advent of Code puzzles by iteratively debugging executions of generated unit tests (intentionally not competing on the global leaderboard).

And even for this, there's actually only a SINGLE place in the whole "agent" where the models themselves actually make a "decision" on what step to take next, and that's simply deciding whether to refactor the generated unit tests or the generated solution based on the given error message from a prior failure.