Microsoft Copilot Studio just changed how AI agents get built for business processes. In early September 2026, Microsoft announced that the GitHub Copilot harness — a new engine for powering agents inside Copilot Studio — is now generally available. Instead of following a fixed script, agents built on this harness can reason through a goal, decide their own steps, and adjust course when something doesn’t go as planned. For anyone who builds, manages, or simply relies on workplace automation, this is a meaningful upgrade to what a “Copilot agent” can actually do on its own.
What Changed?
Copilot Studio has long let makers build agents using a “standard harness,” which follows predefined topics and conversation paths — reliable for FAQs and structured requests, but limited when a task doesn’t fit a predictable pattern. According to Microsoft’s official Copilot Studio blog post from September 2, 2026, the GitHub Copilot harness is now available as a second, more capable foundation builders can choose when creating an agent.
Per Microsoft Learn’s harness documentation, this harness sits between the AI model and the agent, controlling when the model is called, what information it receives, and which tools it can use — the layer that lets an agent plan multi-step work instead of answering one prompt at a time. It also adds native support for creating and reasoning over Word, Excel, PowerPoint, and PDF files, persistent memory across a conversation, and connections to external tools through MCP (Model Context Protocol) servers and other agents.
What Can You Do With It?
The practical difference shows up in how an agent handles a messy, real-world task. Microsoft’s own example is an accounts payable agent: instead of a rigid workflow, it can read an incoming invoice, match it against a purchase order, notice missing information, go find it, and route anything unusual to a human — adjusting its steps as it goes rather than failing the moment something is out of order.
The same flexibility applies elsewhere: an onboarding agent pulling together documents and equipment requests across departments, a research agent drafting a PowerPoint deck from source material, or a support agent checking several systems before answering a customer. Skills and “connected agents” let one agent delegate parts of a job to another specialized agent, so builders can assemble larger workflows from smaller, reusable pieces instead of one monolithic bot.
Why It Matters
For individual users, this mostly means agents that feel less brittle — fewer dead ends where a bot simply doesn’t understand an unusual request. For business and SMB leaders, the bigger deal is automation coverage: processes too irregular for rule-based bots (exception handling, multi-system lookups, document-heavy approvals) are now realistic candidates for an agent, without a developer hand-coding every branch of logic.
It also changes how IT teams think about governance. Because reasoning agents decide their own steps, Microsoft is pairing this release with expanded admin controls — including a way to flag “high-risk” agents for closer monitoring — a sign that more autonomy needs more oversight, not less.
How to Use It
Builders don’t need custom code to try this. In practical terms, getting started looks like this:
- Open Microsoft Copilot Studio and start creating a new agent (or open an existing one).
- When choosing the agent’s foundation, select the GitHub Copilot harness instead of the standard harness.
- Describe the agent’s goal in plain language, then connect the knowledge sources, connectors, or MCP servers it needs to do the job.
- Add skills or connected agents for any sub-tasks that make sense to delegate rather than handle directly.
- Test the agent in the built-in test pane with a realistic scenario, including messy or incomplete inputs, to see how it adapts.
- Publish the agent for internal teams or external customers once you’re satisfied with how it reasons through edge cases.
Because outcomes depend heavily on how clearly a goal and its guardrails are defined, it’s worth testing with intentionally imperfect inputs before rolling an agent out broadly.
Availability
The GitHub Copilot harness is generally available now in Copilot Studio, and agents built with it can be published to internal teams or external customers. Unlike the standard harness, which can draw on Microsoft 365 Copilot licensing in some scenarios, agents on the GitHub Copilot harness are billed entirely through consumption-based Copilot Credits — covering building, testing, evaluating, and running the agent, regardless of the maker’s own license. Microsoft’s Power Platform admin documentation notes that developer and trial environments move to this usage-based billing on September 1, 2026, and recommends admins identify agents using this harness, review consumption history, and set monthly limits before scaling up. Check your Power Platform admin center for current environment and capacity requirements before a broad rollout.
The Takeaway
The GitHub Copilot harness moves Copilot Studio agents from “follow the script” to “figure out the steps” — a real capability jump for automating the kind of irregular, judgment-heavy work that rule-based bots have always struggled with. That flexibility is genuinely useful, but it comes with a real cost model attached, so the smart move for any team piloting this is to test it on a bounded process first, watch the Copilot Credit consumption closely, and expand from there once both the agent’s behavior and its running cost are well understood.
Original Source:
Microsoft — What’s new in Copilot Studio, August/September 2026: GitHub Copilot harness, agent skills, and richer context