Skip to content

Which Harness Runs Inside Copilot, and Do You Need the Best One?

There are four harnesses around Copilot Studio. Microsoft calls the Microsoft Copilot harness (formerly the GitHub Copilot harness) the most capable option. For coding, Claude Code and Codex are ahead, I think. For teams that work with documents, dashboards, and meetings, "the best harness" is the wrong question.

By , Microsoft MVP Published Updated

A man holds up a climbing harness with plum-colored webbing and opens his other hand as if asking which one; on the laptop beside him are a gear, a speech bubble and a code symbol, with the speech bubble framed.
AI-generated image

A harness is the runtime between your agent and the language model. There are four around Copilot Studio, and which one is best depends on what you need it for.

This started with a comment under one of my TikToks saying Microsoft’s harness delivers low-quality AI. Here’s my full answer, with sources.

What does a harness actually do?

The language model supplies the reasoning. The harness decides when to call the model, what context it gets, how its answer is interpreted, and which tools run next. That’s how Microsoft describes it in the Copilot Studio documentation (as of October 2026): the model provides the capability, and the harness equips and directs it.

Which harnesses does Copilot Studio have?

There are four different ones. Microsoft Learn’s overview describes three that every agent and workflow in Copilot Studio runs on: the Microsoft Copilot harness, the standard harness, and the customer conversation harness. The fourth, the Copilot Chat harness, shows up in the admin center as a value in the Harness column.

Harness Built for Billing
Microsoft Copilot harness (formerly the GitHub Copilot harness) Multistep business processes: breaks a goal into steps on its own, calls tools, and adjusts when a step fails. Creates and edits Word, Excel, PowerPoint, and PDF files, supports skills and memory (preview), and runs in an isolated sandbox Copilot Credits
Standard harness Rule-based agents and fixed flows: you define the topics, prompts, and paths, and the agent responds predictably Copilot Studio’s existing licensing model, details in the credits article
Customer conversation harness Agents that serve customers over chat or phone: a primary agent with specialized sub-agents that can hand off to a human. Published as chat on a website or with a phone number; supports skills and memory Copilot Credits, usage-based
Copilot Chat harness Extending Microsoft Copilot Chat with your organization’s knowledge, for internal teams only Usage-based, or included in the Copilot license

The four aren’t the same thing. The ones most often mixed up are the customer conversation harness and the Copilot Chat harness: one serves customers over chat and phone, the other extends Copilot Chat for internal teams. For the Copilot Chat harness, the Learn overview pages linked below (retrieved October 31, 2026) only list the value in the Harness column. The table reflects what I read on October 4, so check it again before you decide.

Microsoft renamed the harness on October 30, 2026; before that, it was the GitHub Copilot harness. Many Microsoft Learn pages and the “GitHub Copilot” filter value in the admin center still use the old name (retrieved October 31, 2026). If you search the documentation, you’ll often find it under “GitHub Copilot harness.”

According to Microsoft, neither the standard nor the Microsoft Copilot harness is better across the board. For short, well-defined tasks the standard harness is the better fit, and the Microsoft Copilot harness can burn extra time and credits there without improving the result. It’s still maturing: memory is currently a preview, so check up front that the features and channels you need are available. A working agent on the standard harness doesn’t need to move just because a newer harness exists.

You pick the harness when you create an agent. According to Microsoft, an agent built on the Microsoft Copilot harness can’t be moved to the standard harness, and the reverse is true as well. For how billing works on the standard harness, see Copilot Studio Credits: Understand Them Before They Blow Your Budget. For the Microsoft Copilot harness, Microsoft explains it on the page Manage costs for agents powered by the GitHub Copilot harness. For the customer conversation harness there’s Agents powered by the Customer Conversation harness overview.

Is the Microsoft Copilot harness the same as the GitHub Copilot service?

No. Microsoft says explicitly (still under the name GitHub Copilot harness) that the harness is a Copilot Studio framework for authoring and orchestration, not the GitHub Copilot service. It shares underlying technology and SDK components with GitHub Copilot, but customer data isn’t sent to the GitHub Copilot service. Copilot Studio’s existing commitments on privacy, security, compliance, and data residency still apply.

So Microsoft uses GitHub Copilot technology inside its own framework. Microsoft’s announcement also says that Code in the new Copilot app runs on the same technology as GitHub Copilot. I cover that in Copilot Pricing Changes 2026.

What do Microsoft’s own measurements show?

In November 2026, Microsoft published a paper by the Copilot Studio team that runs both harnesses on the same evaluation tasks with the same models and also analyzes production data from real agents. The difference is architectural. The Standard harness splits every turn into several separate model calls (recognize intent, plan a response, ask for missing input) and hands control back to you as soon as a step fails. The Microsoft Copilot harness lets one model work in a loop until the task is done or its budget of turns, time, or tokens runs out. It runs in an isolated container with a command line and a file system, and according to Microsoft, credentials stay outside with Copilot Studio.

Metric GPT-5.5 Chat: Standard → Microsoft Copilot Opus 4.8: Standard → Microsoft Copilot
Eval pass rate 67.2% → 78.0% 67.0% → 89.1%
Model calls per task 2.6 → 4.3 2.3 → 4.8
Cost-weighted input tokens 15% fewer 21% fewer
Latency per model call 10% lower 36% lower

So the new harness calls the model more often and burns two to 2.6 times as many tokens overall. Yet the weighted total goes down, because each step of the loop just appends to the same context and the model reads the start of it from the prompt cache: the paper reports cache hit rates of 83 to 88% instead of 39 to 45%, and cached tokens count for less. The paper deliberately leaves prices out, so it doesn’t tell you what this means for your Copilot Credits. Only each single call gets faster: because there are more calls, a task often runs longer overall, minutes to hours according to the paper, compared with seconds per turn. The gain on the reasoning model Opus 4.8 is about twice the gain on GPT-5.5 Chat. Microsoft picked both models because they were available on both harnesses at the time; the new harness now focuses on reasoning models and the Standard harness on chat models.

The gap is widest when a tool call fails. In production data, meaning real agents doing different tasks, the Microsoft Copilot harness keeps going in 84 to 91% of those cases, the Standard harness in 7% at most. In roughly two out of five cases, the same tool later succeeds within the same task. The paper shows why that matters: a task with ten tool calls and a 5% failure rate per call fails somewhere about 40% of the time.

Microsoft also tunes the tools themselves. For the Excel Online (Business) connector, the success rate rose from 85.6% to 96.3% over the summer of 2026, with no change to anyone’s agent: the agent now checks which tables a workbook has before asking for tables that don’t exist. On a suite of 20 business process automation test cases, the Microsoft Copilot harness completed 65% of tasks against 36% for an unnamed leading agent platform. Microsoft itself calls that a snapshot, not a ranking.

These are the vendor’s own measurements on its own tasks, not independently verified. The paper stresses that both harnesses stay in the product and offers a rule of thumb: if the task is complex, involves a lot of data, needs code, should end in a presentation, document, or spreadsheet, or runs without a person present, the Microsoft Copilot harness is the right tool.

How does GitHub Copilot compare with Claude Code?

For pure coding, I think Claude Code and Codex are ahead. One benchmark by two Microsoft authors compares Claude Code and GitHub Copilot head to head: BC-Bench, published August 21, 2026. It consists of 101 tasks from Business Central, which means ERP software development in the AL language. Each task ran five times, both harnesses used the same model, Claude Opus 4.6, and the job was fixing bugs:

Metric Claude Code GitHub Copilot
Tasks resolved, mean 68.5% 65.1%
Resolved in all five runs 49.5% 50.5%
Time per run 284 s 314 s

On average, Claude Code comes out about three percentage points ahead. When it comes to solving the same task reliably in all five runs, GitHub Copilot is one point ahead. The authors write that the difference doesn’t reach the usual significance threshold (p = 0.075) and that the choice of model seems to matter more than the choice between these two harnesses. They also note that this rests on only four matched comparisons.

Read this as one data point, not a verdict: one benchmark, one domain, and the authors work at Microsoft. BC-Bench measures the GitHub Copilot and Claude Code harnesses, not the Microsoft Copilot harness inside Copilot Studio. It says nothing proven about that one. Microsoft only states that it shares technology with GitHub Copilot. Microsoft’s measurements above compare two very different architectures, so their numbers can’t be compared directly with BC-Bench.

What do you need a harness for in practice?

In my experience, most people who use Copilot aren’t programmers. They want to create PDFs and presentations, build Excel analyses and dashboards, evaluate meetings, and write emails. They don’t need a coding specialist for that. They need a harness that handles files, runs inside their own Microsoft environment, and is something IT can take responsibility for.

Then there’s what companies expect from Copilot. What I see is that they want everything in the cloud, inside their own Microsoft context, properly governed, with logs, retention, and resilience, and not running locally on a single machine.

It used to be said that the OpenAI models in Copilot weren’t good enough. From what I’ve seen, Anthropic’s models are in Copilot now too, and new models show up there shortly after release. I think the criticism is shifting from the model to the harness.

I use the Microsoft Copilot harness every day, for coding tasks and for productivity tasks, and I run it in production in projects with large companies.

My take: convenience wins in the end

Anyone who already works with Claude Code or Codex knows from experience what’s possible there. For the broad mass of users, I see the law of inertia: before someone installs and sets up a local harness and local MCP servers (connectors that give a model access to tools), they click in Copilot. Maybe the harness there isn’t the perfect one, maybe it costs more credits, maybe it comes with restrictions. If it’s simpler, more available, and faster, it wins anyway.

So for me the question isn’t whether it’s the best harness on the market. It’s whether it works for the people who actually use it. For a rollout to a large number of users, it’s usable and it brings productivity. It isn’t slop.

How do you find out which harness runs in your tenant?

In the Power Platform admin center, go to Manage and Copilot Studio and look at the Harness column in the agent inventory. For each agent, it shows whether it runs on GitHub Copilot, Standard, or Copilot Chat; that’s how Learn names the values (retrieved October 31, 2026). The GitHub Copilot value belongs to the Microsoft Copilot harness. Learn lists no separate value there for the customer conversation harness, so check in your own tenant. Filter for GitHub Copilot: according to Microsoft, only those agents bill every activity on a usage basis.

If you want to work out which agent belongs on which harness for a rollout, that’s what the coaching for IT and admins is for.

Sources

As of October 4, 2026; names and harness overview checked October 31, 2026; Microsoft’s measurements added November 12, 2026. Microsoft tends to change harnesses, benchmarks, and billing quickly.