Key takeaways
- Build simple agents with Agent Builder in Microsoft 365 Copilot, and richer ones with actions and triggers in Copilot Studio.
- An agent is instructions plus knowledge plus actions; the model decides which to use for each message.
- Agents can read PDFs and SharePoint content, call APIs through connectors, send email and run on schedules in Copilot Studio.
- For coding, GitHub Copilot's coding agent and agent mode are the right tools, not Microsoft 365 agents.
- Share agents with specific people or groups, and keep their knowledge sources within what those people may see.
To build a Copilot agent, open Microsoft 365 Copilot and use Agent Builder for a simple agent made of instructions and knowledge, or use Copilot Studio when the agent needs actions, API calls, triggers or multi-step workflows. An agent combines instructions, knowledge sources and actions, and a language model decides how to use them for each request. This page shows how to create one and how it works, then answers what agents can do: read PDFs and SharePoint lists, call APIs, send email, run on a schedule, call other agents and be shared, plus which agent to use for coding.
How do I create a Copilot agent?
In Microsoft 365 Copilot, choose to create an agent, describe what it should do in plain language, add knowledge such as SharePoint sites or files, test it in the preview pane, and share it. For actions and workflows, build it in Copilot Studio instead.
Also asked as: how to copilot agent · how copilot agent works
A first agent in under an hour:
- Purpose: one narrow job, such as answering questions about your department's policies.
- Instructions: who it serves, what it should and should not answer, tone, and when to hand over to a person.
- Knowledge: the SharePoint site or documents that hold the answers.
- Starter prompts: three example questions users can click.
- Test: ask twenty real questions, including tricky ones, and fix instructions where answers go wrong.
Start narrow; a focused agent that answers well beats a broad one that guesses.
How does a Copilot agent work?
When a user sends a message, the agent's orchestrator reads its instructions, decides whether to search its knowledge sources, call an action or reply directly, runs those steps, and has the language model compose an answer, citing sources where it used them.
Knowledge searches retrieve relevant passages, respecting each user's permissions, which is retrieval-augmented generation. Actions are tools with descriptions, such as "create a ticket" or "look up an order", and the model picks them based on those descriptions, so clear names and descriptions matter. In Copilot Studio you can also define topics, fixed conversation paths for cases that must follow exact steps. Every conversation can be reviewed in analytics and transcripts, which is how you improve the agent after launch.
Which Copilot agent is best for coding?
For coding, use GitHub Copilot: agent mode in VS Code and other editors for interactive multi-file changes, and the Copilot coding agent on GitHub for issues you hand off and review as pull requests. Microsoft 365 agents are built for work content, not codebases.
GitHub Copilot's agents can read the repository, run commands and tests, and iterate on failures. Give them well-scoped tasks with acceptance criteria, keep a .github/copilot-instructions.md file with build and style rules, and review every change through normal code review. For comparison, other strong coding agents include Claude Code, OpenAI's Codex and Google's Gemini CLI; try one real task in each if your team is choosing.
Can a Copilot agent read a PDF?
Yes. Agents can use PDFs as knowledge, either uploaded directly in Agent Builder or Copilot Studio, or stored in a SharePoint site or OneDrive folder the agent searches. They answer from the text and cite the document.
Results are best with text-based PDFs; scanned PDFs need text recognition first, so run OCR on them before adding them. Large manuals answer better when split into sensible sections with clear headings. Keep one current version of each document, because outdated copies lead to conflicting answers. In my OCR & Speech Workspace, page-batched OCR and page citations made answers on scanned documents checkable, which is the same principle: clean text in, citations out.
Can a Copilot agent read a SharePoint list?
Yes. Copilot Studio agents can use SharePoint lists as knowledge for questions, and can read and update list items through the SharePoint connector as actions. Results respect the user's permissions on the list.
Use the list as knowledge when users ask questions such as "which laptops are available in the IT pool?". Use connector actions when the agent must do something precise, such as filter items, create a new request or update a status, because actions return exact records rather than a summarised search. Clear column names and consistent values help the model a lot. For large lists, filter in the action rather than asking the model to scan everything.
Can a Copilot agent call an API?
Yes. In Copilot Studio, agents call APIs through prebuilt connectors, custom connectors, Power Automate flows, or REST API actions defined with an OpenAPI description. In Microsoft 365 Copilot, developers add API plugins to declarative agents.
Steps for a custom API: describe the endpoints in an OpenAPI file with clear operation names and descriptions, set up authentication such as OAuth or an API key stored securely, add it as an action, and write instructions on when to use it. Test with realistic requests and check the inputs the model sends, since vague parameter descriptions lead to wrong calls. Validate responses and handle errors, so the agent says "I could not reach the system" instead of inventing an answer.
Can a Copilot agent send an email?
Yes, through an action such as the Office 365 Outlook connector or a Power Automate flow in Copilot Studio. Agents can also draft emails for the user to review and send, which is safer for most cases.
For anything external or important, have the agent draft and ask for confirmation before sending, and log what was sent. Decide whose mailbox sends the email: the signed-in user's or a shared service account, which changes permissions and how replies are handled. Limit who can trigger email actions and what recipients are allowed, to avoid misuse or accidental mass emails. Test with your own address before connecting real users.
Can a Copilot agent be scheduled?
Yes. Copilot Studio supports autonomous agents with triggers, including recurring schedules and events such as a new file, email or record, so an agent can run without a user starting a chat.
Examples: every Monday at 9 a.m., summarise last week's support tickets and post the summary to a Teams channel; when a new invoice lands in a folder, extract the fields and create an approval request. Scheduled agents run under a configured identity, so check what data and actions that identity can reach. Add monitoring: review run history, set alerts for failures, and keep a human approval step before irreversible actions. Power Automate scheduled flows are an alternative when the logic is fixed and needs no AI judgement.
Can a Copilot agent call another agent?
Yes. Copilot Studio supports multi-agent setups, where a main agent hands specific tasks to other agents, such as an IT agent passing a request to a separate HR or procurement agent, each with its own knowledge and actions.
Multi-agent designs work well when domains are clearly separate and owned by different teams. They add complexity: more places for errors, harder debugging and more permissions to manage. Start with one agent and split only when instructions become too long or knowledge sources conflict. Give each agent a clear description of what it handles, because the orchestrator chooses agents by those descriptions, just as it chooses actions. Test hand-offs with tricky requests that sit between two domains.
Can a Copilot agent be shared?
Yes. You can share agents with specific people, security groups or the whole organisation, depending on your role and your organisation's policies. Copilot Studio agents can also be published to channels such as Teams, websites and Microsoft 365 Copilot.
Before sharing, check that everyone who will use the agent may see its knowledge sources; agents respect user permissions for SharePoint content, but uploaded files and some connectors may expose data more widely. Admins can control who may create and share agents and review agents across the organisation. Give the agent an owner who reviews conversations and updates its knowledge, because an unmaintained agent quietly goes out of date.
Sources
GenAI engineer and AI product builder with 2+ years shipping production-grade AI systems: RAG pipelines, fine-tuned LLMs, hybrid retrieval and multi-modal apps across vision, speech and OCR, architected end to end from ingestion to deployment. Leads engineering for SCULT INDIA's 14-member team, founded the 500+ member TechVerse Enclave community and has mentored 200+ students. Three hackathon first prizes: Changethon 2025 (IIT Roorkee), Product Genesis at Vividhotsava 2025 (CSJMU) and BYTEBATTLE (MeetKats).
Open to GenAI roles, hackathon judging, mentorship sessions and guest talks at colleges. For campus sessions the inviting institution covers the visit end to end: travel (a First AC train ticket for journeys up to five hours, a flight for anything longer), accommodation, meals, and pickup and drop; the honorarium and workshop fees are negotiable on WhatsApp or on the call, and requests are welcome from institutions that can cover the visit this way. Email pranjulrathour41@gmail.com.
