Autonomous AI coding strategies

Ideas

Let AI quauntify / qualify how well it progess, or if it got stuck (if the problem seems to difficult / messy). If the raiting is aboce certain level, we should stop and let use know (rather than try ever new iteration cycles)
AI can quite efficiently generate nice progress report in markdown, and also read these progress report after restart. Sometimes it is efficient to save the progress report, to clear the context window, and restart from previous progress report.

Let AI suggest test (unit tests)
suggest AI to first generate some simple clear-cut examples (typical / free of traps)
then suggest to generate unittest for simple edge-cases (expectable traps)
only after these simple test cases are passed successfully, try it on some real-world example suggested by user.