We Are Optimising the Wrong Resource: A Position Paper on Human Attention Architecture

Agents have made output nearly free, so the scarce resource is now the human attention needed to trust and own it. The next architecture layer should measure agent work against a definition of good that a human wrote, route by potential damage rather than agent confidence, and be judged by trusted outcomes per hour of human attention. This position paper argues that agentic AI has moved the bottleneck from producing output to the human attention needed to trust and own it. It reviews why common fixes fail (human-in-the-loop approval, summarisation, confidence-based routing), proposes an output-measurement layer with three mechanisms (a deterministic checker built from human-written tests, KPIs and architectural fitness functions; routing by potential damage rather than agent confidence; protection of remaining attention through periodic review and canaries), and proposes a system-level metric: trusted outcomes per hour of human attention. It closes with practical steps for agent programmes and open questions, including verifiability of non-code outputs and the attention cost of conversational interfaces. Human Attention Architecture is a proposed framing, not an established standard. This is a conceptual position paper; implementation studies and controlled evaluations are still needed.