Explore the GitHub Copilot harness
With the GitHub Copilot harness, you describe what an agent should accomplish, what information and tools it can use, and which limits it must follow. The agent can then plan and adjust its work to reach the goal.
Understand how enhanced orchestration works
When someone submits a request, the enhanced orchestration model decides how the agent should complete it. For a complex request, the agent can:
- Break the request into steps.
- Select the knowledge, tools, skills, or connected agents it needs.
- Complete a step and review the result.
- Continue with the plan, change its approach, or ask for more information.
- Stop when it completes the request or can't continue safely.
You don't need to define every possible path in advance. The agent can adjust its plan when a tool returns an unexpected result or when the request changes. You control this behavior through instructions, permissions, and confirmation requirements.
Some tasks still need a fixed sequence. For example, a transaction might require the same validation and approval steps every time. Use a workflow for these steps. The agent can gather the required information and call the workflow when the transaction is ready.
Identify core agent components
The Build tab organizes an agent around these components:
- Instructions define the agent's role, behavior, and boundaries.
- Knowledge provides information from sources such as websites, files, SharePoint, and business data.
- Tools let the agent retrieve live data or perform actions through connectors, workflows, and other integrations.
- Skills provide reusable instructions for a specific type of task.
- Model provides reasoning and language generation. Select a model based on the work, performance, and cost.
- Connected agents let the agent delegate specialized work to another agent.
- Memory lets the agent retain relevant details from earlier interactions when the feature is enabled.
Together, these components tell the harness what the agent knows, what it can do, and how it should behave.
Define control boundaries
Because the agent can plan and adjust its work, define clear boundaries. Decide which actions it can perform independently and which need confirmation or human review.
A practical control model includes:
- Proceed for low-risk, reversible actions such as retrieving approved information.
- Confirm before actions such as sending a message, creating a request, or updating a record.
- Escalate when an action requires human judgment, formal approval, or privileged access.
Instructions don't replace platform security. Use authentication, permissions, data policies, and connector controls to enforce access.
For more information, see Agents powered by the GitHub Copilot harness overview.