Here is a punchy, structured translation of your breakdown. It uses the dialect of systems theory and computer science typical of HN discussions, making it a perfect fit for a reply to simonw.
This implies a fundamental systems-theory trade-off:
• A fully Autonomous System is one where all necessary information is front-loaded. The first communication is effectively the last communication. For this to work safely, the system’s scope must be strictly bounded (e.g., restricted to a closed world or a tiny sandbox).
• A fully Unbounded System is one where the agent is pre-authorized to take any action in the open world, but no information is front-loaded. Because the environment is dynamic, the missing context, intent, and boundaries must be supplied continuously. This turns the process into communication over time.
This is a strict information-theoretic trade-off. You cannot have a system that is simultaneously fully autonomous and fully unbounded.
To lower "communication over time" to zero (the AI industry's "set-and-forget" marketing promise), you must severely bound the agent's capabilities. If you unbind the agent and let it roam the real world, you must scale up communication over time, turning the human into a constant feedback loop.
The current marketing hype is a logical fallacy: it promises an unbounded agent that requires zero ongoing communication. But you can't engineer your way out of a fundamental trade-off.