Copilot Studio – GitHub Copilot harness GA + licensing

Copilot Studio · Licensing · Governance

The new harness is GA — and your Microsoft 365 Copilot licence no longer covers it.

Copilot Studio now ships three harnesses. Only one of them bills you while you are still building the agent, and it bills every runtime call even for fully licensed Microsoft 365 Copilot users.

HARNESS BUILD TIME RUNTIME — M365 COPILOT USER Copilot Chat harness Customise M365 Copilot with knowledge NOT BILLED INCLUDED IN USL fair usage limits apply Standard harness Topics, flows, rule-based agents NOT BILLED INCLUDED IN USL fair usage limits apply GitHub Copilot harness Agentic, end-to-end business processes USAGE-BASED authoring · evals · preview USAGE-BASED — ALWAYS BILLED Copilot Credits, no USL inclusion Source: Microsoft Copilot Studio Licensing Guide, August 2026.

Two harnesses stay inside the seat licence you already bought. The third one runs on its own meter — during design as well as in production.

Microsoft has been rebuilding Copilot Studio in public for months: a new agentic orchestrator, a streamlined authoring surface, a redesigned workflow canvas. On 3 August 2026 the new workflows designer reached general availability (message centre MC1442234), and the August 2026 Copilot Studio Licensing Guide formalised what runs underneath it.

The interesting part is not the UI. It is one line in the change log on pages 4–6 of that guide: “New Copilot Studio harnesses: Copilot Chat, Standard, GitHub Copilot.”

Skim past it and you will miss the biggest commercial shift in Copilot Studio since messages became Copilot Credits.

Three harnesses, not one product

A harness is Microsoft’s term for the scaffolding that lets an agent plan, reason, hold context and interact with tools, systems or other agents. Copilot Studio now officially ships three of them, and the one you pick determines both what your agent can do and how you pay for it.

Harness What it is for Example
GitHub Copilot Optimise end-to-end business processes with an agentic build experience Accounts payable agent that reads invoices, matches them to purchase orders and routes exceptions for approval
Standard Rule-based conversational agents with predefined topics and flows IT onboarding agent routing laptop requests through approval and provisioning
Copilot Chat Customise Microsoft 365 Copilot with your organisation’s knowledge Employee onboarding agent answering from SharePoint content

The GitHub Copilot harness is the new experience: natural-language authoring, instructions instead of topics, Skills, memory, connected agents, and the Build / Preview / Evaluate / Monitor surface that collapsed nine configuration tabs into four. Microsoft’s own figures for the new orchestrator claim roughly 20 % better evaluation performance with about half the net token usage.

There is no migration path — in either direction. Agents built in the classic experience cannot be converted to the new one, and new-experience agents cannot be moved back. Different architectures, different runtimes. Existing agents keep running; adopting the new harness means a rebuild, not an upgrade.

The rule that changed

Here is the part that will surface in someone’s budget meeting.

Under the Standard and Copilot Chat harnesses the deal has been familiar for a year: employee-facing runtime in Microsoft channels is included in the Microsoft 365 Copilot User Subscription Licence — subject to fair usage — when the agent operates under the authenticated identity of a licensed user. Building an agent by hand costs nothing.

Under the GitHub Copilot harness, that inclusion does not apply.

Microsoft billing matrix comparing Copilot Chat, Standard and GitHub Copilot harnesses across agent creation, agent runtime and supporting features
Microsoft’s own billing matrix: “No changes to Copilot Chat or Standard harness business models. Introducing usage-based billing for GitHub Copilot harness across agent creation and runtime.” Source: Microsoft Copilot Studio licensing documentation, August 2026.
Copilot ChatStandardGitHub Copilot
Manual (non-LLM) authoringNot billedNot billedNot billed
Natural-language authoringN/ANot billedUsage-based
EvaluationsN/ANot billedUsage-based
Test pane / Preview tabNot billedNot billedUsage-based
Runtime — users with M365 CopilotNot billed (fair use)Not billed (fair use)Usage-based
Runtime — users without M365 CopilotPublished rate cardPublished rate cardUsage-based
Runtime — external channelsN/APublished rate cardUsage-based

Two consequences fall out of that table.

Build time is now billable

Natural-language authoring, evaluations and even the preview/test pane consume Copilot Credits on the new harness. Manual, non-LLM configuration still does not — but the entire point of the new experience is that you don’t configure it manually. A maker iterating on instructions all afternoon is now generating a bill, and the evaluation suites Microsoft has spent a year telling us to run are metered too.

The Microsoft 365 Copilot licence no longer covers runtime

A Microsoft 365 Copilot USL still grants the Copilot Studio Author role, and publishing an agent itself consumes no credits. But because runtime does, you need at least one Copilot Credit purchasing option configured before you can publish a GitHub Copilot harness agent at all. Publishing Standard or Copilot Chat agents to Microsoft channels as a USL holder still requires nothing.

File one related change alongside it: since July 2026, Work IQ API usage by Copilot Studio agents is also excluded from the Microsoft 365 Copilot USL.

The seat licence now buys you the right to build. It no longer buys you the right to run.

How usage-based billing is calculated

The price of each task is assembled from four inputs, all converging on Copilot Credits.

Diagram showing models, context, tools and runtime feeding into Copilot Credits, with one credit equal to one US cent on the pay-as-you-go plan
Models, retrieved organisational context, tool calls and orchestration runtime all meter into Copilot Credits. Source: Microsoft Copilot Studio licensing documentation, August 2026.
  • Models — the AI model chosen for the agent
  • Context — organisational context retrieved to ground the response
  • Tools — actions invoked through flows, connectors, APIs and tools
  • Runtime — orchestration that runs multi-step and long-running work

On the pay-as-you-go plan, 1 Copilot Credit = $0.01. That single line is what turns the four abstract inputs above into a number a CFO can read.

What a single task actually costs

Microsoft buckets agent tasks into three shapes, with credit estimates attached.

Light, medium and heavy agent scenarios with example prompts and estimated Copilot Credit ranges
Light, medium and heavy agent scenarios with estimated credit consumption per task. Source: Microsoft Copilot Studio licensing documentation, August 2026.
Light task100–300credits · ≈ $1–$3
Medium task300–500credits · ≈ $3–$5
Heavy task> 500credits · > $5

A light task is a handful of knowledge sources and one short deliverable — check the latest daily-ops mail, verify a workbook, post a status summary. A medium task references many sources across timeframes and produces two or more deliverables. A heavy task is a multi-year compliance audit that recomputes data, correlates permits and logs, writes a full report and distributes it across SharePoint, Teams and email.

Now do the arithmetic that matters. One Copilot Studio capacity pack costs $200 per month for 25,000 credits. That is roughly 80–250 light tasks, or 50–80 medium tasks — per month, for the whole pack. Compare that with the Standard harness, where a generative answer bills in single-digit credits and tenant-graph-grounded retrieval in the low tens.

The unit economics of an agentic task sit one to two orders of magnitude above a chat turn. That is not a defect — you are paying for multi-step work that previously required a developer — but it retires the mental model of “we bought a pack, we’re covered.” Unused credits still do not roll over, and exceeding purchased capacity still triggers technical enforcement up to service denial.

What is still not billed

The guide is equally clear about what stays outside the meter, on every harness:

  • Standard, premium and custom connectors
  • On-premises and cloud data transfer for Power Platform connectors
  • Storing structured agent data in Dataverse (defaults: 15 GB database, 20 GB file, 2 GB log)
  • Managed Environments, for Copilot Studio-related features
  • Management from the Power Platform admin center
  • Publishing an agent

So the billable surface is narrow and specific: LLM-powered maker experiences and agent runtime on the GitHub Copilot harness. Everything structural around it remains included.

What to do this month

  1. Decide harness by harness, not tenant-wide. A SharePoint Q&A bot on the Copilot Chat harness is still effectively free for licensed users. Do not rebuild it on the new harness because the demo looked good.
  2. Configure a credit purchasing option before makers hit the wall. Pay-as-you-go at $0.01/credit is the lowest-friction start; capacity packs at $200 per 25,000 credits suit a predictable baseline; the Copilot Credit Pre-Purchase Plan (5 % at 300,000 credits, up to 20 % at 300 million) and the Microsoft Agent Pre-Purchase Plan (ACUs spanning Copilot Studio and Foundry, 5–15 %) suit committed spend.
  3. Run the Copilot Credit Estimator before approving a build, and again after evaluations. It is guidance, not a quote.
  4. Govern who may author. The Copilot Studio Author role now gates people who can generate spend simply by iterating. Set the security group in the Power Platform admin center deliberately.
  5. Review PPAC → Licensing → Copilot Studio → Summary weekly, not monthly. With no rollover and monthly enforcement, surprises only ever run in one direction.
  6. Budget evaluation cycles as a line item. They are metered now — and skipping them to save credits is the worst available trade.
If you cannot say which harness an agent runs on, you cannot say what it costs.

Microsoft has effectively split Copilot Studio into a bundled tier and a metered tier. Conversational agents that extend Microsoft 365 stay inside the seat licence you already bought. Agentic agents that genuinely do multi-step work run on their own meter — and that meter runs during design, evaluation, testing and production alike.

That is a defensible position; long-horizon agentic work costs what it costs, and GitHub Copilot went through the same reckoning with token-based billing in June. But it means “can we afford this agent?” now has to be asked before the maker opens the canvas, not after the invoice arrives.

Sources

Written by Lukáš Oplt, edited with AI. Pricing and terms are taken from the August 2026 licensing guide and are subject to change without notice — verify against the current guide before committing a budget.

Comments

Popular posts from this blog

Your Copilot Studio Agent May Be Acting as You — Not the User

Let AI build the flow. Never outsource the blast radius