新しい Copilot Studio が GA したので「ハーネス」とクレジット課金体系を改めて整理してみるWith the GA of the new Copilot Studio, this article systematically explains the…

Microsoft has brought the new version of Copilot Studio to general availability, and the change has pushed two practical questions to the front of the discussion for anyone planning to deploy agents: how the platform actually runs those agents, and how usage is billed. A recently published Qiita article takes on both, organizing what is currently known about the agent execution foundation described as the "harness" and the credit-based billing structure that governs consumption.

The piece is explicitly framed as a reference compiled from public information available as of August 2026, citing the Copilot Studio official blog, Microsoft Learn, and the Copilot Credits Guide (August 2026 edition). Its goal is to help teams evaluating adoption or licensing understand the moving parts, not to announce new capabilities. The author also cautions that these details are a point-in-time snapshot and are subject to change, recommending that readers verify the latest official information before making contractual or deployment decisions.

Copilot Studio itself is Microsoft's low-code environment for building conversational and, increasingly, autonomous agents that connect to enterprise data through connectors and can take actions across Microsoft 365 and other systems. It sits within a broader stack that includes Microsoft 365 Copilot, the Power Platform, and Azure-based AI services, so decisions made in Copilot Studio often intersect with existing licensing and governance arrangements.

The technical heart of the article is the "harness," which it presents as the execution layer responsible for actually running an agent. Conceptually, this appears to cover how an agent interprets instructions, invokes tools and connectors, reasons over the context it is given, and produces a response or completes a task. Because each of these steps can consume compute, the way the harness executes work is directly tied to what an organization ultimately pays.

That connection is why the article pairs the execution model with billing. Microsoft has been moving Copilot Studio toward consumption-based pricing, and the credit system is the mechanism for metering that consumption. Rather than charging solely on a flat per-seat basis, a credit model draws down a balance as agents perform work, which means the cost of a deployment is likely to scale with how heavily its agents are used and how complex their tasks are. Understanding which activities draw credits, and at what rate, is therefore central to forecasting spend.

For teams that previously worked with earlier Copilot Studio billing, where usage was often expressed in terms of "messages," the shift