253: The Adversarial Agents Are Arguing Again

What if the best way to get good work out of AI is to stop being nice to it? Adam and Tim have both landed on the same uncomfortable discovery: when you pit AI agents against each other, with fake points, opposing incentives, and competing models, the output gets dramatically better than anything a single polite prompt can produce.Adam's bug-hunting pipeline hands fake rewards to sycophantic agents and then throws the scores in the trash. Tim made Claude and ChatGPT argue for twelve rounds straight until they both said "ship it".Follow the show and be sure to join the discussion on Discord! Our website is workingcode.dev and we're @workingcode.dev on Bluesky. New episodes drop weekly on Thursday.And, if you're feeling the love, support us on Patreon.With audio editing and engineering by ZCross Media.Full show notes and transcript here.