What is Copilot Studio? Agents, harnesses, channels and cost

What is Copilot Studio? Microsoft's low-code tool for building AI agents and workflows: what you build, where agents run, how it differs from Copilot, and who pays.

Copilot StudioQuick reference12 min read

Checked against Microsoft Learn 22 Oct 2026 · 17 sourcesPosted by the Maker DeskHow we write guides

Copilot Studio is Microsoft's low-code tool for building AI agents and workflows. An agent follows instructions you write, answers from knowledge you connect (SharePoint, websites, files, Dataverse) and uses tools to act (connectors, workflows, REST APIs, MCP servers). You publish it to Teams, Microsoft Copilot, a website or other channels. People who chat with a published agent need no licence of their own: the organisation pays for activity in Copilot Credits.

Note

Checked against Microsoft Learn on 22 October 2026, and against Microsoft's Copilot Studio pricing page.

What goes into a Copilot Studio agent, where it runs, and how it relates to Copilot and Agent Builder.

What you build

Copilot Studio has three kinds of building block, and they can call each other: a workflow can call an agent for one step, and an agent can run a workflow as a tool.

Agents hold a conversation or carry out a task. You describe the agent in plain language or configure it by hand, test it, then publish it.
Workflows are automations built in a drag-and-drop designer, where a step can reason with AI as well as follow fixed logic. They're the newer flow format.
Agent flows are the established flow format, close to Power Automate cloud flows but native to Copilot Studio. They're deterministic: the same input takes the same path. An agent can call one as a tool, and every action they run spends Copilot Credits.

Building apps in Copilot Studio is also in preview.

What an agent is made of

Part	What it does
Instructions	Who the agent is, its tone, what it may and may not do
Knowledge	Where answers come from: SharePoint, public websites, uploaded files, Dataverse tables, Copilot (Graph) connectors and, on the standard harness only, Azure AI Search
Tools	How it acts: Power Platform connectors, custom connectors, workflows or agent flows, REST APIs, MCP servers, prompts (standard harness), computer use (standard harness)
Topics	Scripted conversation paths with questions and conditions (standard harness; the GitHub Copilot harness replaces them with hooks)
Connected agents	Other agents it hands work to: Copilot Studio agents in the same environment, Microsoft Foundry agents, agents that speak the A2A protocol, and Microsoft 365 Agents SDK agents
Triggers	Events that start the agent without anyone typing

Our guides go deeper on choosing knowledge sources and connecting an MCP server.

Conversational and autonomous agents

A conversational agent waits for someone to ask. An autonomous agent starts itself: an event trigger, such as a new SharePoint item, a new email or a schedule, sends it a payload and it follows its instructions. Microsoft's examples include triaging incoming requests and processing updates in the background.

Two points to know before you build one:

On the standard harness, event triggers authenticate only with the maker's credentials. Anyone who can use the published agent may reach data through that connection, so Copilot Studio warns you before publishing. Admins can block event triggers with data policies.
On the GitHub Copilot harness, triggers are in preview. After publishing you choose Activate this agent for me (it runs with your connections) or Share with others (each person sets up their own connections).

On the GitHub Copilot harness, triggers sit in the configuration panel on the agent's Build tab. What the panel shows, as of October 2026:

Model: the AI model behind the agent's reasoning.
Triggers: "Choose the event or schedule that starts your agent". Each trigger is a chip, such as a Teams chat, a new email or a schedule, and the + beside the heading adds another.
Skills, Tools and Knowledge: structured behaviours, the systems the agent can act on, and the sources it answers from.
Memory: whether the agent remembers details from earlier interactions.

The Build tab of an IT Help Desk Agent in Copilot Studio, built on the GitHub Copilot harness. Along the top are the agent's name; the Build, Preview, Evaluate and Monitor tabs, with Monitor greyed out; save, share, feedback and more icons; and a Publish button. On the left are the agent's title and its Instructions, which tell it to answer from the IT Help Desk FAQ and Device Policy, ask one question at a time, check for an open ticket, create a ticket when a person is needed, reply to new emails in the help desk mailbox, and post a weekly summary of open tickets to the IT team in Teams. On the right is the configuration panel. Model is set to Claude Opus 4.7. Triggers, described as Choose the event or schedule that starts your agent, lists five triggers, each with a remove button: Chat with the agent in Teams, Agent @mentioned in Teams, Chat with the agent in Microsoft Copilot, New email in Outlook, and Every 7 days. Skills lists Ticket Triage, Laptop Setup Guide and Password Reset Steps. Tools lists Teams MCP, Help desk tickets in Dataverse and Send email in Outlook. Knowledge lists IT Help Desk FAQ and Device Policy in SharePoint. Connected agents is empty, and Memory is On. — Recreated by LowCodeStacks from Microsoft Learn screenshots (CC BY 4.0): the Build tab of a Copilot Studio agent on the GitHub Copilot harness, with its triggers, interface as of October 2026, with made-up names and instructions.

What people use it for

Microsoft's own examples cover the range: an internal help desk that answers common questions and routes simple requests; an onboarding agent that answers from SharePoint inside Copilot Chat; a customer-service agent on a website or phone line that hands over to a person when it's stuck; and an accounts payable process that reads invoices, matches them to purchase orders and routes exceptions for approval.

The four harnesses

Everything you build runs on a harness: the runtime between your configuration and the model, which decides when to call the model, which tools to use and what to do with the results. You pick it when you create the agent. An agent built on the GitHub Copilot harness can't be moved to the standard harness, or the other way round.

Harness	Best for	Publishes to	Billing
GitHub Copilot	Long, multi-step work that needs reasoning: it plans, picks tools, retries, and creates or edits Word, Excel, PowerPoint and PDF files. Has skills and memory	Internal users or external customers	Copilot Credits, usage-based, including while you build and test
Copilot chat	Adding your organisation's knowledge to Microsoft Copilot Chat	Internal users	Consumption, or included with a Microsoft Copilot licence
Customer Conversation	Customer service over chat or voice, with a primary agent, specialist sub-agents and escalation to a person	Website chat or a phone number	Copilot Credits, usage-based
Standard	Rule-based agents and structured conversations built from topics, plus agent flows	Internal users or external customers	Copilot Studio's capacity model

Despite the name, the GitHub Copilot harness isn't the GitHub Copilot service. Microsoft says it shares technology with GitHub Copilot, but customer data isn't sent to that service, and Copilot Studio's own compliance and data residency commitments still apply.

Warning

A more capable harness isn't automatically better. Microsoft's guidance warns that the GitHub Copilot harness can overwork a simple job, adding latency and Copilot Credits without a better answer. A working standard harness agent doesn't need to move.

Where agents run

You publish from Copilot Studio, then add channels. Which ones you get depends on the harness, and admins can switch channels off for the whole tenant.

Channel	Standard harness	GitHub Copilot harness
Microsoft Teams and Microsoft Copilot	Yes	Yes
Demo website	Yes	Yes
Your own website (embedded)	Yes	Yes, as a web app in an iframe
SharePoint	Yes	Not yet
Mobile or native app	Yes	Not yet
WhatsApp, Facebook	Yes	Not yet
Slack, Telegram, Twilio, email and other Azure Bot Service channels	Yes	Not yet
As an MCP server or an A2A agent for other clients	Not listed	Preview, in early release environments only

The Customer Conversation harness publishes to website chat or a phone number.

By default an agent uses Authenticate with Microsoft, so people in Teams, Power Apps and Microsoft Copilot sign in with Microsoft Entra ID with no setup. No authentication lets anyone with the link chat with it, and the agent then can't use tools with the user's own credentials. Our guide to grounding an agent safely covers why sign-in matters.

Copilot Studio vs Microsoft Copilot and Agent Builder

Three products share the word Copilot, and two of them build agents.

	Microsoft Copilot	Agent Builder (Copilot Studio lite)	Copilot Studio
What it is	The AI assistant people chat with, and Copilot inside Word, Excel, Outlook and Teams. Called Microsoft 365 Copilot until 2026	A quick way to build agents inside the Microsoft 365 Copilot app and Teams	A separate web app at `copilotstudio.microsoft.com`
Who uses it	Everyone	Information workers	Makers and developers
Builds	Nothing; it uses agents	Question-and-answer agents for yourself or a small team, from SharePoint, Copilot connectors and the web	Agents and workflows for a department, the whole organisation or customers
Actions on other systems	Through agents	No: copy the agent to Copilot Studio to add connectors or workflows	Yes: connectors, workflows, REST APIs, MCP, autonomous triggers
Managed in	Microsoft 365 admin center	Microsoft 365 admin center	Power Platform admin center, with environments, data policies and solutions

Microsoft's naming is untidy here. The Agent Builder page on Microsoft Learn sits at a copilot-studio-lite address, release notes call it "Copilot Studio lite", and the pricing page uses both names. They're the same thing.

Use Agent Builder when one person or a small team needs answers from content they can already see. Move to Copilot Studio when the agent needs a wider audience, actions on other systems, an external channel, or proper testing and deployment across environments. Copy to Copilot Studio keeps the agent's instructions and configuration, so nothing is lost.

There's also a smaller Copilot Studio for Teams plan inside some Microsoft 365 subscriptions. It builds classic agents that publish only to Teams and store data in Dataverse for Teams, with no generative orchestration and no Power Platform connectors.

Weighing Copilot Studio against Microsoft Foundry or Power Automate instead? See which Microsoft agent tool to use.

How it's paid for

Is it free? Partly. The maker's Copilot Studio user licence is free once the tenant has a credit pack, and a trial lets anyone build and test, but not publish.
The organisation pays for use, in Copilot Credits. On Microsoft's pricing page in October 2026, a pack of 25,000 credits a month costs $200; pay-as-you-go through an Azure subscription and a pre-purchase plan (up to 20% off) are the alternatives.
A Microsoft Copilot licence lets that person build and use internal agents on the Copilot chat and standard harnesses at no extra cost. External channels, people without a Copilot licence, and the GitHub Copilot harness need standalone Copilot Studio credits.

What each answer, tool call and flow action costs, and what happens when credits run out, is in Copilot Studio licensing and Copilot Credits. How the same credits pay for other Microsoft AI features is in Copilot Credits explained, and the seat licences are in Microsoft Copilot licensing explained.

Ready to try it? Your first Copilot Studio agent walks through instructions, knowledge, testing and a small first release.

Sources

Was this helpful?

Comments

No comments yet. Share what worked for you.