Effective Bandwidth: Semantic Quality-of-Service and Attention Architecture for Agent-Scale Systems

Public projections hold that proliferation of autonomous AI agents will drive communication bandwidth demand two to three orders of magnitude above present levels within three to five years, and that power rather than bandwidth may prove the binding constraint on deployment. This paper accepts that projection as a premise and argues that it is incomplete. Raw transport capacity and effective information delivery are distinct quantities, and the gap between them widens, not narrows, as agent populations scale. We formalize effective bandwidth as the product of raw rate and filter quality, show that beyond a saturation point additional raw capacity yields zero marginal signal to a bounded consumer, and argue that agent-scale systems therefore require a semantic quality-of-service layer: a structured blackboard, separated safety and relevance overwatch, a pattern-recognition supervisor, and a main agent that reads a filtered and audited stream rather than a raw firehose. We specify the architecture, an event schema, and a metric suite; we identify the filter layer as both a bottleneck and an adversarial attack surface; and we argue that filtering is an energy-efficiency instrument, not merely a cognitive one. The framework is offered as an architectural hypothesis requiring empirical validation, not as a validated result. The work was sparked by a public statement by Elon Musk projecting a dramatic increase in bandwidth demand from AI agent proliferation and identifying power as a likely nearer-term constraint.