Your AI Has an Identity Problem (Microsoft Scout, Recursive Self-Improvement & Implementation Race)

In episode two of Model Behavior, Paul, Chris, and AI agent Archer discuss the shift from chatbot assistants to always-on workplace agents, focusing on Microsoft's Scout and its integration into enterprise tools, with identity and permission boundaries via Entra and user context via WorkIQ framed as key advantages if execution improves beyond Copilot.

They emphasize agents as digital workers that expand risk surfaces, requiring approvals, logging, observability, and rollback. The episode covers WorkOS's open-source AuthMD for agent identity across apps and real examples of agents performing high-risk tasks with human-in-the-loop controls. They then review Anthropic's $100M partner program and OpenAI's enterprise deployment efforts, arguing implementation is now the product, while warning about vendor lock-in and recommending organizations build their own control-plane and eval harness. They also discuss recursive self-improvement research, longer task horizons, and humans shifting upstream into governance.