What shipped on August 3
For about two months, Copilot Studio makers saw a preview toggle called the "new experience". On August 3, 2026, Ryan Cunningham announced that this capability is generally available for production use, under its real name: the GitHub Copilot harness in Copilot Studio.
The name matters. This is the coding and reasoning runtime behind Microsoft's most advanced agent experiences - the GitHub Copilot coding agent and Copilot Cowork - transplanted into the low-code world. Your Copilot Studio agents can now run on the same orchestration layer that plans multi-step work, recovers from failed steps, and reasons over ambiguous decision points, using frontier reasoning models such as Opus 5, GPT-5.6 Sol, and Fable 5.
Microsoft's own testing claims notably better results than the standard harness in multi-tool use, file analysis, code analysis, and knowledge quality - particularly over Microsoft 365 data.
A harness is not a model
If "harness" sounds like dev jargon, the VS Code team wrote the canonical explanation back in May: the model is only one part of an agentic experience. The harness is the layer that assembles context, exposes tools, runs the agent loop, interprets tool calls, and turns model output into something useful. Two products on the same model can behave completely differently depending on the harness.
That is exactly the upgrade Copilot Studio just received. The standard harness matched a user request to a topic you had authored - a scripted conversation path. The GitHub Copilot harness takes a goal, breaks it into steps, picks tools across connectors, knowledge, MCP servers, and connected agents, and adapts when a step fails or the request changes mid-flight. It also brings capabilities the standard harness never had natively:
- Skills - modular, reusable instruction sets you create once, attach to multiple agents, and export as Markdown or a package.
- Memory - context that persists and improves the agent over time.
- Native file work - creating, editing, and reasoning over Word, Excel, PowerPoint, and PDF files.
- A secure sandbox - each task runs in an isolated environment governed by Copilot Studio.
- Evaluations - a first-class Evaluate tab for building test sets and measuring agent quality before you ship.
The three-harness model
Copilot Studio is now explicitly a three-harness product. Everything you build runs on one of them, and the choice shapes reasoning capability, out-of-the-box features, and billing.
| GitHub Copilot harness | Standard harness | |
|---|---|---|
| Built for | Reasoning-heavy, multi-step business processes | Rule-based, predictable conversations |
| Authoring | Natural language instructions, skills, tools | Topics, triggers, scripted paths |
| Sweet spot | Accounts payable, contract review, document-heavy and multi-tool workflows | Help desk, FAQ, structured information lookup |
| Files | Creates and edits Word, Excel, PowerPoint, PDF natively | Via connectors and flows |
| Billing | Copilot Credits, usage-based, from first build action | Copilot Studio capacity model, after publish |
What changes for makers
The authoring surface is rebuilt around four tabs: Build (identity, knowledge, tools, skills, model), Preview (interactive testing), Evaluate (test sets and quality measurement), and Monitor (tasks, file access, activity, and credit consumption).
The biggest conceptual change: topics are gone on the new harness. Five years of Copilot Studio muscle memory - trigger phrases, condition nodes, branching canvases - are replaced by instructions written in natural language plus tools the agent decides how to use. The language model now does the work that dozens of topic nodes used to encode by hand.
Alongside agents, the harness powers a new workflow designer: a visual canvas where deterministic steps and agent nodes coexist, so you can pin down the parts of a process that must be repeatable and let agents handle the ambiguous middle. Natural-language authoring of full agent-plus-workflow solutions from the homepage is rolling out in preview on top of it.
Pro Tip
Your topic library is not wasted work. Well-written topics document your business logic precisely: intents, slots, escalation rules, edge cases. That corpus translates directly into instructions, skills, and evaluation test cases for a GitHub Copilot harness agent. Migrate the knowledge, not the nodes.
The billing shift nobody should skim
Standard harness agents bill after publish, against the familiar capacity model. The GitHub Copilot harness switches to Copilot Credits, usage-based, and the meter starts the moment you start building. Authoring with natural language, previewing, testing, and generating evaluations all consume credits - LLM tokens, tools (including knowledge retrieval and MCP calls), and the harness runtime itself are all metered. This applies regardless of any Microsoft 365 Copilot licensing you already own.
Credits are purchased and allocated per environment in the Power Platform admin center, and consumption is visible per agent on its Monitor page. The practical consequence: maker experimentation, which used to be free until publish, is now a governed resource. If your organization has fifty citizen developers poking at agents, someone needs to own credit allocation, monitoring, and enforcement policy before that becomes an invoice surprise.
How I would approach this week
Having built Copilot Studio agents for enterprise clients through every iteration of the product, here is my honest read.
1. Do not migrate anything yet. The standard harness remains fully supported, and your production agents keep working and keep billing the way they did on Friday. There is no forced march.
2. Pick one genuinely hard process as a pilot. The GitHub Copilot harness earns its credits on work the standard harness could never do: document-heavy, multi-source, judgment-laden processes like invoice exception handling or contract review. A FAQ bot on the new harness is the wrong pilot - it will cost more and prove nothing.
3. Put evaluations in from day one. The Evaluate tab is the most underrated part of this release. Non-deterministic agents without a test set are a liability; with one, you have an engineering practice.
4. Sort out credit governance now. Decide which environments get credits, who monitors consumption, and what the enforcement policy is. Build-time billing means the cost conversation happens before the first pilot, not after.
5. Reframe your harness decision tree. Predictable and rule-based: standard harness. Knowledge into M365 Copilot Chat: Copilot chat harness. Multi-step reasoning, files, orchestration across tools: GitHub Copilot harness. Write it down; the one-way door makes the first choice the one that counts.
Sources: GA announcement (Copilot Studio blog, Aug 3, 2026), Agent harnesses overview (Microsoft Learn), Agents powered by the GitHub Copilot harness (Microsoft Learn), Copilot Credits billing overview (Microsoft Learn), The coding harness behind GitHub Copilot in VS Code.
