Build Your First Agent
This walkthrough creates a simple knowledge-and-action agent using Copilot Studio's web experience.
Sign in at copilotstudio.microsoft.com with an account that has Copilot Studio access (a trial works fine).
1. Create the agent
Choose Create → New agent. You can describe what you want in natural language and let Copilot Studio scaffold a starting point, or skip straight to manual configuration. For a first agent, try describing it directly: "An agent that answers employee questions about our expense policy and can submit a reimbursement request."
2. Write instructions
Open the agent's Instructions and refine what was scaffolded (or write from scratch): define its purpose, tone, and any boundaries — for example, "only answer questions based on the connected expense policy knowledge source; don't make up amounts or thresholds."
3. Connect a knowledge source
Add a knowledge source — for a first test, a single uploaded document (your expense policy) or a SharePoint site works well. This is what lets the agent answer specific questions accurately instead of guessing.
4. Add a tool
Add a tool/action the agent can take — for example, a Power Automate flow that creates a draft reimbursement request record. Give the tool a clear description of when it should be used, since the orchestrator relies on that description to decide whether it's relevant to a given request.
5. Test in the chat pane
Use the built-in test chat to ask the agent a few realistic questions: a direct policy question, a request to submit a reimbursement, and something outside its scope entirely (to confirm it doesn't answer questions it shouldn't). Watch which knowledge source or tool the orchestrator picks for each, and adjust your instructions or tool descriptions if it picks wrong.
6. Publish
Once testing looks good, publish the agent to your chosen channel (Teams, a website, or Microsoft 365 Copilot). Review the analytics pane after it's been used for a while to see which questions it struggled with.
What's next
From here, reasonable next steps are: adding a second knowledge source, adding a topic for one tightly-controlled scenario you don't want left to generative orchestration, and — once you have a second agent — connecting the two with agent-to-agent communication so they can delegate work to each other.