Guarded autopilot
Guarded autopilot is an execution policy for autonomous coding agents: tasks run without a human watching, but only inside hard guardrails: an allowlist of who may author them, an injection scan before any prompt is assembled, live budget enforcement, and a run prompt that forbids irreversible actions so the agent stops and describes them instead of executing.
Full autonomy and full supervision are both wrong defaults. Full autonomy ships mistakes at 3am; full supervision means the agent saves no time. Guarded autopilot is the middle that works: the machine's owner decides whose tasks their machine will run, every task's text is scanned for prompt-injection shapes before execution, budgets are enforced by killing the process rather than by warning, and the run prompt forbids pushing or deploying so the agent stops and describes such an action instead of taking it.
In Aethereum the enforcement is local: the guardrails run on the machine executing the run, not as advisory text in a prompt. Nothing executes anywhere until that machine's owner has explicitly enrolled it.