The Cognitive Firm: Authority, Independence, and Control in Organizations of Humans and AI Agents

Meta-agents increasingly allocate work, evaluate subordinate agents, and revise the procedures under which they operate. We argue that permission to propose an evaluator revision should be distinguished from authority to make it binding. Freezing every evaluator preserves defects; unrestricted revision lets the producer change the terms of success. We develop the Cognitive Firm as an organizational response, drawing on decision control, normative multi-agent systems, and contestable AI. The central distinction is between repairing an evaluator under an existing policy, amending that policy, and changing who may approve either. An authority graph records rights over proposals, evidence, ratification, and appeal. A scheduling example traces the different evidence and approvals these changes require. We propose comparisons that isolate revision rights while counting useful outputs, false releases, false rejections, and oversight costs. When agents manage people, affected workers also need standing to challenge the governing policy. This is a position paper; the comparative benefits of the proposed arrangement remain to be tested.