On August 3, Microsoft made the GitHub Copilot harness in Copilot Studio generally available. Most of the coverage framed it as a more powerful way to build agents, which is true, and which buries the lead. The lead is one word, harness.
A harness is not a model and not a feature, Microsoft is using this word for a reason. The harness in Copilot Studio is the thing you now choose before you build. And the GitHub Copilot harness has one defining trait. Instead of following the steps you laid out, it plans its own. That single trait is why it is powerful.
It is also why it shows up on your bill in a way the old ways never did. Powerful, while expensive, for the same reason.
The layer between you and the model
We need to start with what a harness even is, because Microsoft is now asking you to pick one. This used to be a more simple setting you could change as needed.
Microsoft’s own definition is clean. You design the agent, the model supplies the reasoning, and the harness is the runtime that sits between them. The harness decides when to call the model, what to send it, how to read what comes back, and which tools to invoke. The model thinks. The harness acts.
The practical shift is that the harness, not the model, is now the primary unit you select. The model rides underneath it.
Three harnesses, and a word the industry already uses
Microsoft of course did not invent the concept of harnesses in AI. People building agents have been using it for a while. Anthropic describes the system that powers Claude Code as an agent harness, and points out that the same harness can drive agents that have nothing to do with coding. Claude Code, Codex, and Cursor are all language models wrapped in a harness that gives them tools, memory, and an execution loop the base model does not have on its own. My own OpenClaw instance, Falcon, sits in that same lineage. What changed is that Microsoft turned the harness into something you pick from a menu.
Copilot Studio now has three.
The standard harness is the one most existing agents already run on. You define the topics, the prompts, and the paths, and it follows them. Predictable, rule-based, and well understood. We’ve been using this “harness” for years, well before anyone was thinking of it as a harness.
The Copilot chat harness extends Microsoft 365 Copilot Chat with your organization’s knowledge, so employees get grounded answers without leaving the tools they already work in. This is the “harness” that was released a few years ago as a part of generative AI orchestration.
The GitHub Copilot harness is the new one, and changes our agent build thinking. It does not follow a script. You give it a goal, it breaks that goal into steps, calls the tools it needs, and adjusts when something fails. And your billing moves to the moment you start building an agent using this new harness, for every turn.
That last difference is the one that really matters. On the standard harness, you control each step. On the GitHub Copilot harness, it controls the flow. Everything else follows from that, including your costs.
What it buys you, and what everyone else already built
Here is what that trait buys you.
The GitHub Copilot harness brings the coding and reasoning capabilities behind Microsoft’s most advanced agent experiences, Copilot Cowork and the GitHub Copilot coding agent, into Copilot Studio. It plans and recovers, calls tools across connectors, MCP, and other agents, holds memory, and natively creates and edits Word, Excel, PowerPoint, and PDF files, with each task running in a sandbox. Microsoft’s own example is an accounts payable process, where an agent reads invoices, matches them to purchase orders, and routes the exceptions for approval.
That is not a chatbot. That is a tailored, targeted business process running end to end.
Every major vendor seems to have shipped a harness. Salesforce has Agentforce, Google has the Gemini Enterprise Agent Platform, Amazon has Bedrock AgentCore. The existence of an agentic runtime is not the news, and neither is low-code agent building. Agentforce is low-code. Google’s Agent Studio is low-code. AWS sits at the code-first end, a bring-your-own-framework runtime built for engineering teams.
So Microsoft’s actual move is more than just “agents for makers.” They took their frontier coding-agent runtime, the same family behind Cowork and their coding agent, and dropped it behind our low-code canvas wired directly into Microsoft 365 and Power Platform.
The canvas remains the one we have been using since PVA’s. The engine behind this new harness, and where it plugs in, is definitely new.
The catch, there is always a catch
Which brings us back to the harness’s new trait. The GitHub Copilot harness plans its own steps. That is what makes it capable. It is also what makes it cost money in a way the standard harness does not.
It is the headline gotcha of the whole release.
On the standard harness, billing starts after you publish. On the GitHub Copilot harness, the meter starts when you start building!
Creating an agent with natural language, previewing it, testing it, and running evaluations all consume credits.
And you are not only paying for model tokens. Credits are used to cover the tokens, the tools and knowledge and MCP calls, and the harness itself.
Wait, what?
This is a big deal, a huge shift in considering agents, agent building, and ROI. The old motion was build freely, then pay once it is live, often wrapped into a prepaid subscription. The new motion is that iteration itself is metered. The autonomy you were sold at the top is the autonomy you fund at the bottom. For the entire end to end lifecycle of your agent.
The full economics of this, what it does to forecasting and to governance, is its own conversation.
What this means
If we step back, the direction of AI is more clear.
The model is one dropdown . You pick it from a menu, and you can swap it. Easy.
The harness is the product.
That matches a position I have held for a while: the raw capability of the model stopped being the thing that decides value. Where the value lives, and now where the cost lives, is the layer that deploys the model into real work. Microsoft just made that layer the thing you shop for and the thing you pay for. Same layer, both sides of the ledger.
The bottom line
So here is what actually happened. Microsoft did not just ship a more powerful agent builder. It made the harness the product, put a frontier coding agent behind a low-code canvas, and moved the meter to the moment you start building.
The capability is real. The harness is the product. The billing is the story.
Pick the harness that fits the job. Just know that on this one, thinking is billable.
Speak Your Mind