The Agent Is You, at a Distance: How human intent gets lost

Chat is simple: prompt input gives you llm output to review. Agents are not. Agents create side effects before you see them. That changes the boundary: agents are now acting with a partial model of human intent and whatever authority we give them, but how well and at what distance does the intent decay?

This talk starts with a real failure from my persistent personal agent, Hermes. We built a sophisticated heartbeat to be proactively useful. It ran 1,404 times, made 1,004 model decisions, sent 43 messages, and took one useful quiet action. Every component worked perfectly. Yet the system was still wrong.

From that case I derive a delegated-intent loop and the concept of action distance: how far an agent acts before human intent must be synchronized again.