SudoAnirudh/Self_Correcting_Agent ? reverse-engineered prompt

Build me a Python project for an autonomous research agent that can take a hard web research question, break it into smaller steps, gather facts, and write a final structured answer. I want it to keep a simple working memory of what it has learned so far, check each step to see if the result looks useful or off track, and if something fails or looks wrong it should try to recover by retrying, rewording the search, or adjusting the plan. It should also have a basic baseline mode that just pushes forward without self correction so I can compare the two.

Please include support for flaky tool behavior like timeouts, bad outputs, and partial results so the self correcting loop can be tested properly. Add automatic model fallback if the main provider is unavailable. I also want an evaluation runner for a small set of research tasks, a report that compares baseline versus self correcting performance, unit tests, sample logs, and a simple HTML log viewer to inspect what happened during each run. Look up current docs online if you need to.