Define an agent's purpose and scope
Before you create an agent, define the job it owns and the decisions it doesn't own. A narrow, explicit scope gives you a practical basis for writing instructions, selecting knowledge, and testing behavior.
Describe the agent's job
A useful purpose statement identifies the audience, the task, and the intended outcome. For example:
Help Contoso employees understand company benefit programs by answering questions from approved benefits information and directing personal eligibility decisions to Human Resources.
This purpose statement is more actionable than Help employees with benefits. It identifies who the agent serves, what information it provides, which sources it should use, and where its responsibility ends.
Define these elements before authoring:
- Audience: The people who interact with the agent.
- User goals: The questions or tasks the agent supports.
- Expected outcomes: What a successful interaction produces.
- Available information: The sources the agent can use.
- Out-of-scope requests: Questions or tasks the agent should decline or redirect.
- Escalation destination: The person, team, or process that handles exceptions.
Separate information from decisions
An information agent can explain approved content without making decisions that require authority, personal records, or professional judgment. For a benefits agent, explaining plan terms is different from deciding whether a specific employee qualifies for coverage.
Use boundaries to distinguish these responsibilities. A boundary should state both what the agent must not do and what it should do instead. For example, Don't determine an employee's eligibility. Explain the available criteria and direct the employee to Human Resources for a decision. The second sentence gives the agent a useful alternative instead of only prohibiting behavior.
Create the agent
On the Copilot Studio Home page, the Agent tile creates an agent powered by the GitHub Copilot harness. You can also select Agents and then New Agent. The agent opens on the Build tab, where you configure its identity, instructions, knowledge, and other components.
Enter the agent's name before the first save. The name used at that point determines its schema name, which becomes read-only. If your organization uses solutions, also choose the correct solution before saving the agent.
For more information, see Build a new agent.