Something along the lines of "Read the wikipedia page of the day. Verify every single link and the context matter against all files in this computer. Then trace their correlations to each other, showing which link corresponded to most files by subject matter, after that is done, verify your work by doing the same from different starting points. We expect similar results. After 100 rounds of that, it should be good. Then you should create a DB to store all that data (only after you ran the full 100 verificaritions yourself) and reverify every field against the pages and the files"

That should keep it going for an hour. Turn on fast mode and auto mode (if using claude) for extra costs.

Every page and file will increase its context, burning tokens

If you need some plausible deniability about it being real work and not just obviously you running up costs:

Feed it a bunch of work-related documentation and then have it do a bunch of reviews of the content on that documentation.

you would be mostly burning your own money if you did this, so I wouldn't recommend it (depends on how the agent is priced)