Looking for extreme / impossible tasks to properly stress-test my agent.I can’t trust my own judgment anymore

​ I built a fully autonomous custom agent architecture. I give it a task and completely leave it alone. It can run for hours or days (longest continuous run so far was 3 weeks) without any intervention. It handles its own errors, decides what tools and steps it needs, and keeps going. Some of the things it has already done in my own tests: - Continuous run of 3 weeks with zero human intervention - Wrote an 800-page manuscript by itself with research for old books - In roughly 9 out of 10 long-running tasks the context window does not fill, even after days of continuous work I know these are big claims and hard to believe. I’m stating them on purpose, because if I post something more modest, people will only send average tasks. Here’s the real reason I’m posting this: I can no longer be objective. It’s very possible that I’m stuck in my own loop / illusion and that the agent only looks good because the tasks I gave it were ones I subconsciously knew it could handle. I need external, extreme, even impossible tasks to see the truth. I don’t just want to know if it finishes the task. I want to see: - Does it get stuck or loop? - Does it block / crash? - How does it actually handle truly hard or adversarial situations? What I will publish: Only the final, unedited output of the agent on GitHub. No traces, no reasoning steps, no tool calls, no intermediate data (proprietary). Here I posible to be a deal breaker for many, but at the moment is not possible. I’m taking 5 most extreme tasks, no matter how crazy or adversarial they are. If you have something that has broken other agents or frameworks before, or something you consider nearly impossible for current agents, drop it here. I need the reality check. Thank you too everyone who will decide to take the time, read and give me a task.