Agents Are Not Users: A Reference Architecture for Governed Agent Actions in Third-Party Products

Artificial intelligence agents are increasingly executing actions inside third-party products, yet existing authorization models were designed primarily for human users or machine-to-machine integrations. As agents begin updating records, sending emails, triggering workflows, and performing other side-effecting operations, organizations require governance mechanisms that go beyond traditional API authentication. This paper presents a reference architecture for governed agent actions in third-party products. We introduce an action-centric model based on declarative manifests, explicit authorization boundaries, scoped delegation, approval workflows for irreversible actions, audit receipts, and verifiable execution semantics. Rather than treating agents as privileged API clients, the architecture separates reasoning from execution and places policy enforcement at a dedicated action gateway. The paper formalizes the architectural components, trust assumptions, security considerations, and design principles underlying this approach, providing a conceptual foundation for building secure, auditable, and governable AI agent integrations.