← All insights

News analysisLens: United States5 min read

Copilot Studio after Ignite 2026: the new harness bills from the first test. Plan the budget before the build

The Microsoft Copilot harness is now generally available in Copilot Studio, with better task results and a different meter: credits start when you start building. What changed and what to set first.

An old timing instrument with a round dial in a wooden case, in La Madre duotone, beside the words The clock starts early
Photo: MKFI (Wikimedia Commons, Public domain)

The most consequential Copilot Studio change from Ignite 2026 is not on the keynote slides. It is one sentence in Microsoft’s billing documentation: agents built on the new Microsoft Copilot harness consume Copilot Credits “from the moment you start building.” Previewing an agent, testing it and generating evaluations all count. On the older harness, now called the Standard harness, billing starts after you publish.

That sentence changes how a Copilot Studio project should be budgeted, and who needs to be in the room before the first prototype.

What Microsoft shipped

Copilot Studio had no section of its own in the Book of News. Its updates are spread across a Dynamics 365 post by Jeff Teper, who leads the product, a Copilot Studio blog post and the Learn documentation. The ones that matter for teams building agents:

  • The Microsoft Copilot harness is generally available in Copilot Studio. It is the same runtime Microsoft uses for Copilot itself, previously called the GitHub Copilot harness inside Copilot Studio. Microsoft’s internal evaluation against the Standard harness, on equivalent tasks and models, reports knowledge answer quality up more than 20 points, multitool task success from 87.0% to 98.9% and 45% fewer input tokens per task. These are Microsoft’s numbers from successful evaluation runs, and Microsoft itself says they vary by day, build, model and configuration.
  • Local development. Developers can build agents and workflows locally, keep the artifacts in GitHub, and publish to Copilot Studio from the command line. Command-line tools and a Copilot Studio MCP server are in public preview.
  • Agent sharing is generally available, so several makers can co-author one agent.
  • Computer use through Windows 365 for Agents is available as a tool in the Standard harness, billed in Copilot Credits, and is due in preview on the Copilot harness in December.
  • Agent Accounts, which give agents memory and triggers so they can act as persistent participants in a process, enter public preview in December 2026.
  • Copilot Studio spend appears in Agent 365 Cost Management, next to Cowork, Copilot Managed Runtime and Dynamics 365 agents, with spending policies, limits and alerts.
Two harnesses, two meters
In Copilot StudioStandard harnessMicrosoft Copilot harness
StatusGenerally available, the existing runtimeGenerally available since Ignite 2026
When credits startAfter the agent is publishedFrom the first build, preview, test or evaluation
What is meteredPer feature: answers, actions, grounding, flow actions, AI toolsModel tokens, tools including knowledge and MCP, and the harness itself
Computer use via Windows 365 for AgentsAvailablePreview expected in December
Based on Microsoft Learn billing documentation and Microsoft's Ignite 2026 posts, read on November 20, 2026.

Why the billing change matters more than the benchmark

A better harness is good news, and 45% fewer input tokens per task should lower the cost of each run. But the cost profile of a project also moves. On the Standard harness, a team could iterate for weeks in the test pane at no metered cost and only start paying when real users arrived. On the Copilot harness, iteration itself is consumption. A team that runs an evaluation set of 200 questions after every instruction change is now spending credits on quality work, which is exactly the work you want it to keep doing.

So the budget conversation moves to the start. Three things are worth settling before a maker opens the editor:

Which environment pays for development. Copilot Credits are allocated per environment in the Power Platform admin center. A development environment with its own allocation, or its own pay-as-you-go billing plan, keeps experiments from drawing down the capacity that production agents depend on. On prepaid capacity, Microsoft blocks agents when the tenant reaches 125% of what it bought; a burst of testing should never be what takes a live agent offline.

Which harness each agent should use. The Copilot harness is the better runtime for multistep work with tools and files, by Microsoft’s own measurements. The Standard harness has a published per-feature rate card that is easier to forecast, and it is where computer use is available today. For a simple FAQ agent with classic and generative answers, the older meter may remain the more predictable choice. That is a decision per agent, not per tenant.

Who sees the spend. Agent 365 Cost Management puts Copilot Studio next to the other usage-based Microsoft AI services. Give the owner of each agent a view of its consumption from the first week, not after the first invoice. We have written about which spending controls actually stop a bill; per-environment allocation is one of the few that does.

Local development is the quieter governance win

Building agents in files, in GitHub, published through a command line, means an agent can finally go through the same review a code change does: a pull request, a reviewer, a history of who changed the instructions and when. Copilot Studio already supports solutions and pipelines, but a large share of agents still live where a maker created them. Microsoft recently closed the shared default environment to its own makers for the same reason: an agent without an owner and a history is hard to defend in an audit.

The CLI path is in preview, so treat it as the direction for professional teams rather than the default for every maker. But if your organization is drafting Copilot Studio standards now, write the rule with it in mind: agents that touch regulated data or act in business systems are built from source control, and makers’ personal environments are for exploration.

What to do this month

If you run Copilot Studio in production, check which harness each agent uses before anyone switches them, and estimate the credit impact of moving the busiest one. If you are starting a project, put the development environment’s credit allocation in the project plan next to the licenses. And if your agents need to operate a desktop application, plan on the Standard harness until the Copilot harness preview arrives.

The product got better for complex agents. The meter got earlier. Both are reasons to treat a Copilot Studio agent as a system with an operating budget, which is how we approach Copilot Studio projects.

Have an AI use case stuck between prototype and production?

Tell us what you’re trying to ship. We’ll reply with honest next steps.

Discuss a use case