This is less an article than a workbench with the tools still out.
I keep circling the same handful of obsessions, and this is where they collided: agents that play, judge, and author without stepping on each other's toes; retrieval and memory that lives for exactly one game and then forgets you on purpose; the stubborn problem of getting a language model to actually reach something instead of just talking; a real neural net running in your own browser tab, no server, quietly measuring you; and underneath all of it a question I can't shake — can two minds, one of them made of silicon, land on the same word without ever agreeing to?
I don't know yet what this becomes. A game, a toy, an essay that wandered off, or a probe for something I haven't named. The honest answer is that I'm finding out by building it, and what you're reading is the workbench, not the finished thing.
But it runs. So for anyone who'd rather play than watch me think out loud, here's how it works: