How to create an agent or workflow in the Copilot Studio GitHub Copilot harness

Microsoft’s quickstart for the GitHub Copilot harness in Copilot Studio is a no-code path. Describe an enterprise process, and Copilot Studio decides whether you get a deterministic workflow, a conversational agent, or both. You test the generated artifacts, then publish by opening the agent or workflow, as the later sections describe. This article covers that preview quickstart and what the harness is. It does not cover GitHub Copilot Autopilot, cloud sessions, or Microsoft Agent Framework harnesses.

What the harness is, and what it is not

The GitHub Copilot harness is a Microsoft Copilot Studio authoring and orchestration framework, not the GitHub Copilot service. It shares underlying technology and SDK components with GitHub Copilot, but customer data is not sent to or processed by the GitHub Copilot service when agents run in Copilot Studio. Existing Copilot Studio privacy, security, compliance, and data residency commitments continue to apply.

The harness is a redesigned authoring and runtime environment. You describe the agent in natural language, and the system generates the configuration, instead of authoring explicit conversation topics, flows, and branching logic. The overview marks that natural-language step as preview. The quickstart calls the feature a production-ready preview subject to supplemental terms of use. The quickstart page is prerelease documentation and is subject to change.

You choose the harness when you create a new agent. Agents on this harness sit alongside standard-harness agents, and an agent created on one harness cannot be transferred to the other.

This harness uses its enhanced orchestration model for every agent. In the standard harness, orchestration behavior can be configured. Microsoft describes deeper reasoning and higher-quality responses than the standard harness, particularly over Microsoft 365 data. Standard-harness agents remain fully supported.

Usage-based billing applies to using, building, testing, and evaluating agents, and those actions might consume Copilot Credits. The natural-language experience might use Anthropic models if they are available in the environment, so contact an administrator.

Create the solution from a description

On the Copilot Studio home page, describe the process, system, or problem in the input box. Be as specific as you like. The quickstart’s example is:

Automate our employee onboarding process: send welcome emails, provision accounts, assign training, notify the manager, and follow up until all steps are complete

More context helps Copilot Studio tailor the result. Select the arrow to submit. A full-page chat opens, and Copilot Studio analyzes the request while it asks clarifying questions. Artifacts take shape in the Artifacts section of the side panel. The result may be a deterministic workflow, a conversational agent, or both.

If the solution needs a system that is not connected, Copilot Studio might prompt you to configure a connection. The documented example is email: if the solution must send mail and no email account is connected, you are prompted to set up permissions with your email provider.

Finished artifacts appear as cards in the chat. Select the Open icon on a card to view or customize it. You can keep editing in the chat, and you can move between the chat and the canvas. Past conversations are on the Chat page and are retained for 28 days. Select one to resume editing or start new artifacts.

The overview’s agent lifecycle is create, build, test, publish, and monitor. Creation starts on the home page when you choose a new agent powered by the GitHub Copilot harness.

What the Build tab covers

The Build tab holds the core components:

Instructions define identity, tone, scope, and behavior.
Knowledge is connected data sources and memory that give the agent context for answers.
Tools let the agent act, such as calling APIs and running flows.
Skills are reusable structured behaviors.
Model is the AI model used for reasoning.
Connected agents delegate specialized tasks to other agents.
Memory, when turned on, lets the agent remember interaction details and reuse them later.

On the agent authoring surface, Preview is the interactive test chat. Evaluate is where you create and run test sets that measure quality. Monitor reviews recent tasks, files the agent accessed, and activity. The Build tab brings authoring into one place so orchestration can interpret instructions, decide when to use knowledge, and decide when to invoke tools. The standard harness instead uses explicit topics, triggers, and branching flows. The quickstart and overview name these areas, including knowledge, but they do not spell out every add-component click. Use the labels on the Build tab for those steps.

Test before you publish

Test, change, and test again before you publish.

If the solution created an agent, open it and select the Preview tab. Ask questions in the test chat. Then open the Build tab and change the instructions. The quickstart suggests a different tone, for example telling the agent to talk to users like Jane Austen, and comparing how responses change. Ask another question, then select Save.

If the solution includes a workflow, open it and edit steps: reorder them, add steps, or remove steps. Select the Run icon in the top right corner of the canvas, follow the prompts, and try different inputs. Review the results and edit again if accuracy or performance is off.

Delete an artifact

On the generated artifact card, select the three dots, then Delete agent or Delete workflow. Verify the name in the confirmation message and select Delete. The artifact leaves the chat context and the designer’s artifact list. A workflow deletion also removes its run history. Deletion is permanent and cannot be undone. While the artifact is saving, publishing, or not yet published, delete is unavailable and a tooltip explains why.

Publish an agent or a workflow

A trial license lets you create agents and test them in the test chat panel. You cannot publish the agent on a trial.

Publishing targets a live or demo website and depends on the platforms and channels you can access. You might need to change authentication for a demo website, depending on who should reach the agent. The quickstart does not list those authentication settings.

To publish an agent to a demo website:

Open the agent and select the chevron next to Publish. The Publish agent dialog opens.
Select Demo website in the list of available channels.
Select Save and publish. Copilot Studio publishes the agent to a demo website, and a URL is displayed.
Copy the URL and send it to others to demonstrate the agent.

To publish a workflow, open it and select the Publish button on the canvas. Post-publish monitoring for workflows is not described on these pages. Comments on the agent authoring experience go through the feedback button in the top toolbar.

What to try next

Ask your administrator whether Anthropic models are available in the environment and whether actions that might consume Copilot Credits are acceptable. Describe one specific process in the home page input box, submit it, and answer the clarifying questions while you watch the Artifacts panel. Test on Preview or with Run, save instruction changes, and publish an agent to a demo website only if you are not limited to a trial license, which can test in chat but cannot publish. If you have not created the agent yet, read the harness overview for the non-transfer rule before you pick a harness.