One pattern I’ve been using in Astra (that I can see @emollick is doing as well) is asking it upfront for hypotheses. Then spinning up dozens to hundreds of agents to go off and research and validate and discover, each with different context windows. If we believe AI will start to have scientific breakthroughs in the next 18 months (which many experts do), then including more hypotheses testing could prove to be more helpful, ie is AI just hammering 1M options and randomly found a good one…