Copilot Agents · 5 min read
Copilot Studio workflows explained: what does what, and where the partner signs off
By James Wilkinson 4 October 2026
In the new Copilot Studio, the agent does the thinking and a workflow makes sure the right person signs off. Get that split right and most of the confusion about the new building blocks goes away.
TL;DR
- The agent makes the judgement calls, a skill holds your firm's way of doing a task, a tool does one action in one system and a workflow runs the steps that must happen the same way every time.
- Partner sign-off belongs in a workflow, not in the agent's instructions. A workflow step is a gate; an instruction is guidance the agent interprets.
- A workflow's route is fixed but any AI step inside it can still word things differently each run, so put AI steps before the human check, never after it.
- Every workflow action draws on capacity or Copilot Credits, so a workflow should earn its length.
Microsoft has added a new way to build agents in Copilot Studio, running on what it calls the GitHub Copilot harness. It comes with four building blocks you will hear about from anyone building agents: agents, skills, tools and workflows. This guide explains each one in plain English, using a job every firm knows: getting an engagement letter from request to partner sign-off.
Checked against Microsoft’s documentation on 4 October 2026. This area changes often, so check the date before relying on the detail.
Four building blocks, one engagement letter
Each building block does one kind of work. Picture a manager asking for an engagement letter for a new limited company client, and follow it through.
The agent makes the judgement calls. It reads the request, works out which services are involved and spots what is missing, such as the year end or whether payroll is in scope. Then it asks the manager rather than guessing. Microsoft describes an agent as something that reasons through a request and decides the best next step.
A skill holds your firm’s way of doing it. How you scope a letter, which schedules go with which services, the wording your risk partner insists on. A skill is written once as a set of instructions, with templates or reference documents attached if needed. Any agent that drafts letters can reuse it, and the agent picks it up when a request matches.
Tools do single actions in your systems. Look up the client record, open last year’s letter in SharePoint, save the new draft. Each tool is one capability, usually a connector to a system you already use.
A workflow runs the steps that must happen the same way every time. Send the draft to the responsible partner, wait for approve or reject, file the approved letter and tell the manager. A workflow can be started by an agent, on a schedule, by an event or by hand.
One point that trips people up: a workflow is added to an agent as a tool. That is why the two terms blur. The difference is size. A tool is one action; a workflow is a chain of them, with rules about the order.
The partner’s decision is the gate: nothing is filed until they approve, and a rejection goes back to the agent with their comment.
Why sign-off belongs in a workflow, not the agent’s instructions
You can write “always get partner approval before sending” into an agent’s instructions, and the agent will usually follow it. Usually isn’t good enough for an engagement letter.
Instructions are guidance the agent interprets each time it runs. A workflow step is a gate. The letter can’t move on until the partner has responded, and their answer decides what happens next.
Workflows include human-in-the-loop steps that pause the process, send the reviewer a request (by email in Outlook, for example) and carry on once they respond. Ask for a comment as well as a yes or no. Then add a step that saves the decision and the comment to the client file, so you can show who approved what and when.
A good rule of thumb: the agent drafts, a person approves.
”Deterministic” needs a footnote
The route through a workflow is fixed, but the words an AI step writes inside it are not.
Microsoft describes workflows as deterministic, meaning the same input produces the same output. That holds for the route: the steps run in the order you set and none get skipped. But workflows can also contain AI steps, such as running a prompt or calling an agent, and those can word things differently from one run to the next.
The practical rule: put AI steps before the human check, never after it. Anything an AI step writes should pass a person before it reaches a client.
What about agent flows and Power Automate?
Neither is going away, and anything that already works doesn’t need rebuilding.
Agent flows are still what Microsoft calls the established flow format in Copilot Studio, and they can include human review steps too. They run on the standard harness, which Microsoft positions for rule-based agents and predictable, repeatable work. Workflows are the automation side of the newer GitHub Copilot harness, built for longer processes that need reasoning across several tools and files.
So the choice is less “old or new” and more “what does this job need”. A well-defined, rule-based process suits the standard harness. A process that needs judgement along the way suits the GitHub Copilot harness. Power Automate is still there for automation that doesn’t involve an agent at all. Our guide to which harness your agent should be built on goes through that choice in more detail.
And no, your data isn’t going to GitHub. Despite the name, Microsoft states the GitHub Copilot harness is not the GitHub Copilot service. Customer data isn’t sent to or processed by that service, and Copilot Studio’s existing privacy, security, compliance and data residency commitments still apply.
What it costs to run
Every step costs something, so a workflow should earn its length.
Workflows draw on your Copilot Studio capacity for each action they run. A ten-step workflow costs more per run than a single tool call, so don’t build a workflow where one action would do.
If your environment’s prepaid capacity runs out, new workflow runs are blocked until more is available. Runs already under way finish, but nothing new starts. Pay-as-you-go billing avoids that stoppage, at a price.
Agents and workflows on the GitHub Copilot harness use Copilot Credits, and Microsoft notes that building and testing can consume credits too, not just live use.
Before approving a project, ask whoever is building it for an estimate of runs per month and what each run uses.
A quick way to decide
Ask one question of each part of the process, and the answer tells you which building block it needs.
| Ask | Use | In the engagement letter |
|---|---|---|
| Does this need a judgement call? | Agent | Working out which services apply and what’s missing |
| Is this how our firm does a particular task? | Skill | Your scoping rules, schedules and standard wording |
| Is it one action in one system? | Tool | Opening last year’s letter in SharePoint |
| Is it several steps that must happen in order, or a sign-off? | Workflow | Partner approval, then filing and notifying the manager |
Where to start
Most firms don’t need all four building blocks on day one. Pick one process where the sign-off matters, map who decides what, then build the smallest thing that does the job.
If you’d like a hand working out where the agent stops and the partner starts in your own processes, book a free 30-minute call. No pitch, no obligation.
Sources checked
Last checked: 4 October 2026.
Related reading
More on Copilot Agents
Common questions