The agent-economy fork: Anthropic puts agents on a meter while Microsoft makes them a governed asset class

Two announcements 48 hours apart frame the operating-model fight for the agent stack. Anthropic is forcing programmatic Claude usage onto its own credit pool with no rollover. Microsoft, at Build 2026, is pitching Agent 365 as the place enterprises register, govern, and police every agent — its own and everyone else's. The pricing layer and the governance layer are both being pulled out of the model.

For most of the last year the dominant agent question was capability: can an autonomous loop actually finish a multi-step task without a human babysitter. As of this week, that has stopped being the interesting question. The interesting question is who owns the meter and who owns the control plane — and two of the three frontier vendors just answered it in opposite directions.

Anthropic confirmed June 15 as the cut-over date for separating Claude Agent SDK, claude -p, Claude Code GitHub Actions, and third-party SDK-built agents from the normal subscription pool. Pro plans get a $20 monthly credit, Max 5x gets $100, Max 20x gets $200, all metered at full public API rates, with no rollover and no automatic overflow unless the developer manually opts in to usage-based billing. The interactive surfaces — Claude.ai chat, Claude Code in the terminal, Claude Cowork — keep drawing from the existing flat subscription. What Anthropic just shipped is a pricing reclassification: agents are no longer a feature of your subscription, they are their own product line.

The same week, Microsoft used its Build 2026 keynote in San Francisco to position Agent 365 as the enterprise control plane for every agent on a managed Windows tenant — its own Copilot agents, Semantic Kernel agents, LangChain agents, and anything else that lands through the new Agent 365 SDK. Agent 365 reached general availability on May 1; Build added the SDK, the Scout always-on governance agent, Agent Orchestrator (preview August), and native Windows integration that ties agent execution into Entra identity, Intune device posture, Defender threat protection, and Purview data classification, with the Windows local-agent surface entering preview in July.

These two moves are not in direct competition, which is exactly what makes the pairing useful. Anthropic owns the model and the SDK; it is monetizing the inference loop that an autonomous agent generates. Microsoft does not really care which model the agent runs on — Agent 365 explicitly supports heterogeneous agent teams under one orchestration plane — it is monetizing the identity, observability, compliance, and seat-count surface around the agent. Read together, they are pricing the two halves of the same trade.

The structural consequence for buyers is that the cost of running a Claude-powered autonomous workflow inside a Microsoft-managed enterprise now decomposes into at least three line items: the Agent SDK credit pool at Anthropic API rates, the Agent 365 governance seat (and whatever Scout/Orchestrator pricing emerges), and the Windows licensing footprint underneath. The subscription-subsidy era — where a $20/month Claude Pro silently powered a developer's nights-and-weekends autonomous coding rig — is closing for the same reason that the unmetered SaaS API era closed: the unit economics of a long-running agent are nothing like the unit economics of a human-paced chat session.

What still has to settle is the third layer: the data and tool plane these agents are calling. Last week's enterprise-data MCP wave (ZoomInfo, Hyland, OtterlyAI all exposed governed MCP endpoints on June 1) is the supply side of the same market — the entities whose data the agents will burn credits reading and whose actions the control plane will police. The agent stack is composing into a recognizable shape: model and SDK billed at the Anthropic meter, governance and identity billed at the Microsoft seat, data and tools billed by whoever owns the MCP endpoint. Anyone selling at a single layer needs a clear answer for who is collecting on the other two.

The New Stack — Anthropic splits billing again: Agent SDK gets separate credit pools → · Microsoft — Build 2026: Securing code, agents, and models across the development lifecycle → · Futurum Group — Microsoft Agent 365 Turns Shadow AI Into a Governed Asset Class →