Automation · 6 min read
Copilot Studio's new engine: what the GitHub Copilot harness changes and what it costs
By James Wilkinson 8 August 2026 Updated 30 October 2026
Copilot Studio's default is now the GitHub Copilot harness, the engine behind Copilot Cowork. What changes, what the credits cover and the checks before you build.
TL;DR
- On 3 August 2026 the GitHub Copilot harness became generally available in Copilot Studio and took over the front page. It is the same engine that powers Copilot Cowork and the GitHub Copilot coding agent, running models including Opus 5, GPT-5.6 Sol and Fable 5. The classic experience remains under other ways to build.
- Building changes with it: skills written in plain language replace topics, an agentic loop plans and checks its own work and a workflow designer handles the deterministic processes, with a human in the loop step as a first-class part of a flow.
- The checks: usage-based billing starts while you build, not when you publish. The Microsoft 365 Copilot licence does not cover this harness. The harness choice is permanent. And agents built during the preview moved to credit billing on 1 September 2026.
Update, October 2026: The GitHub Copilot harness launched on Opus 5, GPT-5.6 Sol and Fable 5. Since then Microsoft has announced Claude Fable 5.1, GPT-6 Astra, Claude Opus 5.5, GPT-6 Sol, GPT-6.1 Sol and Claude Sonnet 5.5 for Copilot Studio, and on 25 September it set out which AI work sits in the Copilot licence and which is paid in Copilot Credits. Agents on this harness remain fully credit billed. For the current models and how each is billed, see our living list of the models behind Copilot.
Open Copilot Studio and the front page has changed. Two cards sit at the top, agent and workflow, and both carry the same label: GitHub Copilot. The classic experience your existing agents were built in is still there, tucked under other ways to build. Microsoft is telling you where it wants you to go.
On 3 August 2026 Microsoft made the GitHub Copilot harness generally available in Copilot Studio and made it the default. When we took a first look at the Workflows Designer in May, the open question was billing. That question now has an answer, and it changes how you should approach the tool.
What a harness is
Every AI agent has two parts. The model is the brain, the thing you talk to in Copilot Chat. It reads a request, thinks and writes an answer. The harness is everything around the brain: the tools it can call, its memory, its file handling and the loop that keeps it working at a task until the task is done. Same model, different harness, very different agent.
Microsoft’s own documentation now describes Copilot Studio as a choice between three harnesses:
- The Copilot Chat harness extends Microsoft 365 Copilot Chat with your organisation’s knowledge, so people get grounded answers without leaving the tools they already use. Fair use is included with the Microsoft 365 Copilot licence.
- The standard harness is the classic Copilot Studio, with topics, triggers and rules you define. Predictable by design, billed on the existing rate card, with fair use for licensed users.
- The GitHub Copilot harness is the new default, and the reason for the word engine rather than feature. It is the same harness that powers Copilot Cowork and the GitHub Copilot coding agent, the one developers use to write code autonomously. It runs the current top-tier models, Opus 5, GPT-5.6 Sol and Fable 5, plans a goal into steps, calls tools and connectors, works natively with Word, Excel, PowerPoint and PDF files, remembers context and retries when a step fails, all inside a sandbox governed by Copilot Studio.
What building looks like now
The new authoring experience has four tabs: build, preview, evaluate and monitor. The first thing existing builders will notice is what is missing. There are no topics.
In their place are skills. A skill is a name, a description and instructions written in plain language, describing what the agent should be able to do and when. The agent works out the rest at run time through its agentic loop: it plans, acts, checks its own work and keeps going until the task is done. It is the same shift we described in our SharePoint post: instructions as assets rather than flowcharts of every possible turn. Writing skills, packaging them with their supporting files and orchestrating several agents is the kind of work a Run plan takes on for firms with agents already running.
The workflow designer sits alongside for the processes where you do not want creativity. Workflows are deterministic: same input, same output. A trigger starts the flow, manual, scheduled or on an event such as an email arriving, and a visual canvas holds connectors, branching, loops, AI steps and rule-based steps together. A human in the loop step is a first-class part of the canvas, pausing the flow for approval before it continues. That review point being a named building block rather than a workaround is exactly the right instinct, and it is how agent work should be run: with agreed places where a person checks.
Natural language authoring, describing a business goal and letting Copilot Studio assemble the agents and workflows, is announced as coming next.
The checks before you build
1. You pay while building, not just when it runs. The classic experience billed you when an agent went live. The GitHub Copilot harness meters maker activities too: every preview test conversation, every prompt where AI writes your instructions and every evaluation run consumes Copilot Credits. Manual configuration is free. It is the AI usage that costs, and because the harness itself runs on AI, testing is AI usage. Budget the build phase the way you would budget the run phase, and set expectations before anyone starts iterating in the preview tab. The habits that keep Cowork credits under control apply here without modification.
2. The licence you already pay for does not cover it. Microsoft 365 Copilot licensed users keep fair use of agents on the Copilot Chat and standard harnesses. The GitHub Copilot harness is usage-based billing for all work, regardless of licensing. If your business case assumed the Copilot licence would absorb agent running costs, it holds for the older harnesses and not for this one.
3. The choice is permanent. An agent created on one harness cannot be transferred to another. Build on the new harness and there is no route back to classic. Build classic, under other ways to build, and there is no route forward. Picking wrong means starting from scratch, which makes the harness decision part of scoping, not something to discover mid-build.
4. The September date if you explored the preview. Microsoft’s Message Center notice MC1446644 says agents and workflows created before 3 August kept their old pricing only until 1 September 2026, then moved to credit billing. Microsoft’s admin documentation adds that developer and trial environments moved to usage-based billing on the same date. If people in your firm experimented during the preview months, check what is still switched on, because the meter is now running.
A better engine on a meter
Is this a money grab? That accusation is doing the rounds, and Microsoft’s Ryan Cunningham has answered it directly: you only pay when you are actually using AI. That is true as far as it goes. What it leaves out is that credits pay for the harness as well as the model. An agentic loop that plans, checks and retries makes many model calls to complete one task, so the meter runs faster than a chat window ever did, and testing sits on the same meter.
Neither framing is wrong. The engine is genuinely better, and better here means more model usage per task, which means more cost. Our honest reading: this is what agent work costs when the pricing stops hiding it. The mistake would be treating the new default as the place for every build, rather than the place for the builds that need it.
What to do with this
Match the harness to the work, because the pricing now punishes getting it wrong. We have since written a full decision guide on which harness each agent belongs on, with the four questions to ask at scoping, and costed a first agent’s whole first year for a 30-person firm, harness choice included. The short version:
- An agent that answers questions from your content inside Copilot Chat belongs on the Copilot Chat harness, covered by the licence you already have.
- A rule-based, repeatable conversation belongs on the standard harness, on the rate card you already know. The Copilot Studio or Power Automate question is unchanged for deterministic processes.
- Save the GitHub Copilot harness for a process that genuinely needs reasoning across steps, files and tools, and treat it as a scoped build: one contained process, agreed human review points and a credit budget for build and test settled before anyone opens the designer. The same budget applies to building and testing on it, which is why a Firmwide install estimates the credits during mapping and sets the limit before anything is built.
That sequencing, from the habits your licence already covers to agents where the work justifies the meter, is the agent journey we build around. If you have a process in mind, which harness it belongs on and what it would cost to run are settled during mapping, at the start of a Firmwide install. Book a free Copilot Readiness Check. You leave knowing where Copilot can and cannot see your data, the two jobs to automate first and a rough install price.
Sources checked
Last checked: 3 October 2026.
- More powerful agents and workflows for autonomous business processes: Introducing a new harness for Copilot Studio
- Choose a harness, Microsoft Copilot Studio
- Access standard harness agents and agent flows
- Microsoft Learn, “Manage costs for agents powered by the GitHub Copilot harness”
- Microsoft Learn, “Agents powered by the GitHub Copilot harness overview”
- Microsoft 365 Message Center notice MC1446644, GitHub Copilot harness general availability (visible to administrators in the Microsoft 365 admin centre)
Related reading
More on Automation
Common questions