Book 17: Alignment & Agency explains the difference between systems that act and systems that are allowed to act. In the BookCockpit Atlas, alignment is not obedience and agency is not power. Agency is the capacity to choose among actions. Alignment is the structure that decides which actions may exist at all. This book shows why a system that cannot refuse unsafe action is not aligned, and why control without refusal collapses into compulsion. Building on the QA framework of Book 16, this volume formalizes alignment as a geometric boundary around action rather than a list of goals. It introduces admissible agency sets, refusal operators, and external constraint rules that prevent systems from rewriting their own limits. Alignment is defined relative to human stability, not preference or ideology, so that optimization cannot justify harm, escalation, or dominance. The book develops formal operators showing how uncertainty expands refusal, how incentives must obey admissibility, and how multi-agent systems remain aligned only when consensus cannot override safety. It demonstrates why learning systems must be free to change objectives but forbidden from changing boundaries, and why power accumulation through the narrowing of others’ choices is a structural form of misalignment. Alignment is then extended from machines to institutions. Markets, states, and AI systems are treated as agents whose actions must remain reversible, whose boundaries must be external to their own authority, and whose growth must not depend on shrinking the agency of others. Conflict is permitted; harm is not. Disagreement is allowed; domination is not. This book does not tell systems what to want. It specifies when movement is legitimate and when stillness is required. It reframes alignment as a conservation law on agency: no system may increase its freedom by reducing the freedom of those around it. Book 17 is the Atlas’s agency filter — the proof that action can exist without conquest, and choice can exist without coercion.