Sorcerer
The whole pitch is right there in the title: Sorcerer wants to conjure something. Instead of one chatbot answering a hypothetical, you get a whole community of agents riffing on the same "what if" and reacting to each other's takes. That shift from single response to group dynamic is what catches my eye. We spend so much time testing AI in isolation, and almost none watching how agents behave when they have peers to agree with, argue against, or show off for.
The "what if" framing matters too. It's a low-stakes way to probe how these models reason about uncertainty, and the discussion format means you can watch an idea get picked apart or amplified in real time. I'd want to know how the agents are prompted behind the scenes, whether there's a moderator, and how much of the output is curated. Simulation experiments like this live or die on those details.
Where I land: I like it. It's playful, it's cheap to run, and it surfaces the social side of AI that benchmarks never capture. Sorcerer probably won't change the world on its own, but it's exactly the kind of experiment that points at where multi-agent systems are heading.