Which Copilot Studio Harness Should Your Agent Be Built On?

Every Copilot Studio build now starts with a question that did not exist in July. The front page shows two cards, agent and workflow, both powered by the GitHub Copilot harness. The classic experience sits below them, under other ways to build. The card you click decides what the agent can do, what it costs and whether the licence you already pay for covers any of it. It also cannot be undone, because an agent can never move to a different harness later.

We covered what changed on 3 August and what it costs when the new harness went live. This piece is the decision that follows: which harness each agent actually belongs on, and how to work that out before anyone starts building.

What a harness actually does

Every agent has two parts. The model is the brain, the thing that reads a request and writes a response. The harness is everything around the brain: the tools it can call, its memory, its file handling and the loop that decides what happens next. Same model, different harness, very different agent.

The difference is easiest to see in how each harness handles a task.

A classic Copilot Studio agent works broadly in one pass. It looks at the request, picks the tools it needs, calls them, hands the results to the model and returns an answer. When that goes well, it is quick and predictable. When a tool fails or the answer comes back half done, that is what you get. The agent has no plan to go back to, so it stops.

An agent on the GitHub Copilot harness works from a list. Give it a task and the harness has the model write a plan, then work through it one step at a time. After each step it checks the result. If a step fails, it retries or finds another route. If the results show more work is needed, it adds steps to its own list. It keeps going until the list is done and the output matches what was asked. If you have watched Copilot Cowork tick off its plan while it works, you have seen this engine, because it is the same one.

That loop is why the output is better. Not because the model got smarter, but because the same models get more attempts at the problem, with a structure that makes them check their own work before handing it back.

Better and dearer are the same fact

The loop is also the cost. Every pass around it, plan, act, check, adjust, is made of model calls, and Copilot Credits meter model usage. One task on the GitHub Copilot harness can involve dozens of calls before the list is done. That is why a generative answer on the standard harness is 2 credits, roughly 2p, while Microsoft’s planning ranges for the new harness start at 100 to 300 credits for a light task run, roughly one to three dollars, and climb past 500 for heavy work.

Understanding that changes how the pricing reads. It is not one harness being reasonable and the other being greedy. The meter is measuring work done. The question for any given agent is not whether the new harness is expensive. It is whether the task is worth the run.

The three harnesses and who pays

Microsoft’s documentation now describes Copilot Studio as a choice between three harnesses, each with its own billing:

• The Copilot Chat harness extends Microsoft 365 Copilot Chat with your organisation’s knowledge. Fair use is included with the Microsoft 365 Copilot licence.
• The standard harness is the classic Copilot Studio, with the topics, triggers and rules you define. It bills on the existing rate card, 1 credit for a classic answer, 2 for a generative answer, 5 for an agent action, and usage by licensed users in employee-facing scenarios is covered by the licence. Billing starts when the agent is published, so building and testing are free.
• The GitHub Copilot harness meters everything for everyone. No licence coverage, ever, and the meter runs from the first prompt while you build, test and evaluate, not just after publish.

Two dates matter alongside that. The new harness became the default on 3 August 2026, and anything built during its preview keeps its old pricing only until 1 September 2026. The full checklist of what changed covers both, and our premium licence guide covers what stays included.

The decision: frequency against value

Put the capability question and the billing question together and the decision becomes a simple trade: how often does this agent run, and what is one run worth?

Take an agent that answers policy and procedure questions for staff, the most common first agent in most firms. It might handle hundreds of questions a week, and each answer saves someone a few minutes of searching. On the standard or Copilot Chat harness, each answer is pennies, and for licensed staff it is covered by the licence entirely. On the GitHub Copilot harness, the same conversation could cost a dollar or more, hundreds of times a week, for answers that did not need a plan, a file or a retry. High frequency, low value per touch: the covered harnesses win, every time.

Now take a month-end task: gather the figures from three systems, reconcile the differences, draft the management pack and route it to a named person for review before anything goes further. It runs once a month. It has steps that depend on each other, files to produce and things that can go wrong part way through. That is the shape of work the agentic loop exists for. If a heavy run costs five dollars and the alternative is an afternoon of someone’s time assembling a draft for the same review, the meter is easy to justify. Low frequency, high value per run: the GitHub Copilot harness earns its cost.

The trap is the mismatch: cheap tasks on the expensive meter. The default cards make that mistake easy, because the most capable harness is now the first one everyone sees. Capable is not the test. Worth it is the test.

Four questions before any build

1. Who uses it, and how often? Licensed staff asking frequent questions point to the covered harnesses. A task that fires occasionally points either way. Anything public-facing is metered whichever harness it is on, so model the volume first.
2. What does one run actually save? Put a time figure on the task as it is done today, with the review step kept in. If one run of the agent saves an hour of assembly and checking still sits with a person, a few dollars per run reads very differently from a few dollars per question.
3. Does the work need the loop, or rules? Variable steps, files, judgement between systems and recovery from surprises need the loop. Same input, same output every time is rules, and rules belong on the standard harness or in a flow. Our Copilot Studio or Power Automate comparison still draws that line for deterministic work.
4. Which budget carries it, and where is the cap? Licence fair use costs nothing extra. Credits need a number: set a per-agent monthly limit in the Power Platform admin centre before the first build session, not after the first surprising bill, and remember the new harness meters building and testing too. The habits that keep credit spend under control apply from day one.

Answer those four and the harness picks itself. And because the choice is permanent, they belong at the scoping stage, before anyone opens the designer, not as a discovery mid-build.

Where this fits

The harness question is now the first fork in every agent build, and getting it right is the difference between an agent the licence quietly covers and an agent with a running cost that needs a business case. That sequencing, matching each piece of work to the cheapest thing that does it well, is how we approach the whole agent journey. If you have a process in mind and want a view on which harness it belongs on, and what it would cost to run, our Copilot Studio agents service starts with exactly that scoping, or book a consultation and we will work through the four questions with you.

Sources checked

Last checked: 10 August 2026.

• Choose a harness, Microsoft Copilot Studio
• Overview of usage-based billing, GitHub Copilot harness
• Billing rates and management, standard harness
• Standard harness licensing
• Microsoft Copilot Credits Guide, August 2026