Artificial intelligence in pharmacy has largely been framed as a source of predictions, alerts, classifications, or generated recommendations. Agentic systems create a different governance problem because they may maintain task state, formulate multistep plans, invoke tools, initiate workflow actions, monitor resulting conditions, and decide whether to revise, stop, or escalate. This article develops a non-empirical governance architecture for situating such capabilities within medication-use work. The proposed Bounded Agentic Pharmacy-Work Theory conceptualizes agentic performance as a relationship among a defined pharmacy-work mandate, authorized context, planning and action functions, permission constraints, consequence monitoring, escalation pathways, and accountable human and organizational control. Four interacting layers are proposed: the pharmacy-work mandate, agentic execution, permission and authority, and human and organizational control. These layers are connected through mandate specification, an action envelope, an authority contract, an escalation contract, and a trace-and-consequence ledger. A plan–act–monitor–escalate cycle is introduced to distinguish generation of a plausible plan from permission to cause a medication-related action. Evaluation must therefore extend beyond model accuracy to plan coherence, tool use, permission adherence, action correctness, reversibility, monitoring sensitivity, escalation performance, human factors, medication-safety consequences, equity, robustness, and local workflow effects. The architecture does not establish that agentic systems improve pharmacy practice or that any proposed permission level is clinically, legally, or regulatorily acceptable. Its original contribution is a testable framework for determining how agentic capability may be bounded, evaluated, supervised, restricted, or withdrawn before it is treated as legitimate pharmacy-work authority.