Caveman and the Empty Room: Designing Agents for an Audience of Zero

One Saturday night, I published a Claude Code skill that made an agent talk like a caveman. Three weeks later, tens of thousands of developers were using it. The joke worked because everybody recognized the voice: polite, over-explained English generated inside loops where nobody was reading it.
Caveman began as output compression. It gave me a larger question: how much of the agent stack still performs for a human reader who has left the room? Coding agents narrate plans they could execute, reload context they already learned, dump tool output into their own memory, and send essays to subagents. Those habits come from chatbots. Harness decides whether they survive.
This talk follows Caveman from a viral skill into a way of thinking about agent design. I’ll use Cavemem, Cavekit, Caveman Code, and real benchmark failures to show what changes when language inside a harness becomes infrastructure. Memory can preserve decisions instead of conversations. Specifications can become work queues. An agent can recognize its own limits and route around them. Humans stay at the boundary, where explanation and control matter.
That is where I think agents are going. Chatbots answer until they get stuck. Agents can rearrange their environment, recruit another capability, and continue. Intelligence people experience will come increasingly from harness around model: what it remembers, what it can call, and what it does when its first approach fails.