Whether we call it Cowork, Code or Codex; whether it is delivered by Anthropic, Microsoft, OpenAI, or another player, the visible product layer itself is becoming increasingly incidental.
Product narratives overlap, compete, and cannibalize one another, creating confusion even at the highest levels of enterprise governance. Despite the noise, a common grammar is taking shape. MCPs, portable skills, agent protocols, reusable contexts—like Google's Open Knowledge Format (OKF) recently proposed—are gradually sketching a shared layer above individual products.
Yet what strikes me most is the "inversion of gravity" this reveals on the enterprise workplace.
For decades, much of the enterprise software revolved around a governed collection of applications. MS Office, SAP, Salesforce—each with its own interface, permissions, lifecycle, and budget. Sold per seat, anchored by heavyweight clients and/or sovereign cloud infrastructure. AI was expected to become another feature embedded inside those systems.
Instead, the opposite seems to be happening!
The emerging center of gravity is an agentic operating system: Chat → Cowork → Agent. Applications increasingly look like disposable satellites orbiting around conversational workflows. None of this replaces the large systems of record, the business applications, or the engineers who build and maintain them—but a new layer is forming above them: a continuum of agentic bridges, loosely held together by a shared grammar of models, protocols, and contexts.
Models themselves have become fluid and interchangeable. GPT-5.5 today, Opus 4.8 tomorrow, DeepSeek 4 the day after (or an enigmatic "Cowork 1" in today's press release about Microsoft Cowork GA)—switched almost casually through the chat UI, often with little visibility into where inference really occurs or how workloads are routed.
Data moves across this emerging layer without friction. Yet many of the assumptions inherited from the previous era—sovereignty, governance, inference residency, availability, pricing stability—still draw their legitimacy from an enterprise narrative forged around stable clouds, applications, and infrastructure. Pricing stability is the first to crack: Microsoft's own announcement of metered, task-based pricing for Cowork reads like a quiet death sentence for the per-seat model.
And yet, the enterprise trust capital accumulated over decades still obscures both shifts at once—architecture quietly recentering, pricing quietly cracking—boiling-frog style, the heat rising well before anyone reacts. What fascinates me is that much of our governance still assumes the old cosmology. The real architecture, however, may look more and more different.
