Copilot Cowork, Anthropic’s Integration, Microsoft’s New Bundle

• Source: Stratechery (Ben Thompson)
• Date: 2026-03-10
• Type: daily-update
• RDCO Relevance: High (Anthropic competitive positioning, agent architecture, enterprise AI)

Three interlinked stories about Microsoft’s AI strategy. First, Microsoft launched Copilot Cowork, built on Anthropic’s Claude Cowork technology, integrated into M365. Key distinction: Cowork runs in the cloud within Microsoft’s security boundary, not locally like Claude Cowork. Microsoft positions this as a feature for enterprise security and governance.

Second, Thompson identifies Anthropic’s moat: the integration of model and harness. Copilot Cowork is not multi-model despite Microsoft’s model-agnostic rhetoric because the harness (the software that orchestrates when and how to use the model) is Anthropic’s differentiation. Microsoft can swap base models for chat, but cannot yet replicate what makes Cowork work without Anthropic.

Third, Microsoft launched the E7 bundle at $99/user/month (65% over E5), bundling Copilot and Agent 365. Thompson predicts Microsoft will progressively move must-have features into E7 to justify the price and offset per-seat revenue decline as AI reduces headcount.

The broader framework: Microsoft is executing “commoditize your complements” across its stack, moving from Azure model hosting to platform-level to application-layer model integration. But Anthropic’s harness integration is the one complement they cannot yet commoditize.

RDCO note: The model+harness moat thesis is directly relevant to our own agent architecture. The value is not in which model you call but in the orchestration layer that knows when and how to call it. See also: Jaya Gupta moat thesis (06-reference/2026-04-10-jaya-gupta-anthropic-moat.md).