Build a Copilot agent without code: from idea to working agent in 7 steps

In our training sessions, this is invariably the moment the room sits up straight: the realisation that you don't have to wait for IT or an external agency to build your first AI agent. With Agent Builder — the lightweight version of Copilot Studio that lives inside Microsoft 365 Copilot Chat — every employee can create their own agent using plain language. In this article we walk through the complete step-by-step plan, including the choices people tend to get stuck on in practice.

First: when is building your own agent worthwhile?

An agent is a specialised version of Copilot: its own instructions, its own knowledge sources, one clear task. It pays off the moment you notice yourself typing the same kind of prompt over and over, or that colleagues keep asking the same questions. Classic first agents: a help desk for the staff handbook, a support agent for common IT questions, a quote checker that runs every proposal past your fixed checklist, or an onboarding buddy for new colleagues. Rule of thumb: recurring work, clear sources, a well-defined topic.

Step 1: open Agent Builder in Copilot Chat

Open Microsoft 365 Copilot Chat and choose "Create agent" in the left-hand panel. That same panel also shows your recent chats and the agents you created earlier or that colleagues have shared with you. Don't see the option? Then your organisation may have disabled the creation of agents — more on that later.

Step 2: describe the purpose of your agent

Describe in plain language what the agent does and which tasks it should take on. For example: "I want an agent that acts as an IT help desk. It helps users set up their computer, resolve network problems and diagnose software and hardware issues." The more concrete the task description, the better the agent. Based on this, Agent Builder suggests a name; you can keep it or adjust it.

Step 3: set the behaviour and the tone

Tell the agent what to emphasise and what to avoid. For example: "Emphasise installation guides and step-by-step instructions. Avoid long walls of text, jargon and assumptions." Also choose a communication style: formal, simple, friendly. This is the step that makes the difference between an agent colleagues enjoy using and one they close again after two questions — so take a moment over it.

Step 4: switch on extra capabilities

Under the "Configure" tab you fine-tune the agent: name, icon and extra features. Two options are worth mentioning. Code Interpreter lets the agent use Python for more complex tasks such as data analysis. Image Generator lets it generate images and diagrams. Only switch on what the task genuinely needs — every extra capability is also an extra way for the agent to wander off course.

Step 5: connect your knowledge sources

This is the most important step. Connect the data sources the agent should draw on: SharePoint sites, specific document libraries or public websites. An agent without its own sources gives generic answers; an agent with the right sources becomes a colleague who knows the handbook by heart. Mind the flip side: the agent can only find what the user is allowed to see, and it will also surface documents that have accidentally been shared too widely. So check the permissions on the sources you connect beforehand.

Step 6: test and publish

Test the agent before publishing in the test window with real questions — not the questions you hope will be asked, but the messy variants colleagues actually type. Also ask something the agent isn't supposed to know: a good agent will then honestly say it can't find the answer in its sources. Happy with it? Click "Create" and your agent is ready to go.

Step 7: share the agent with your team

Through the sharing settings you decide who can use the agent: specific colleagues, a team or the whole organisation. Start small — share it with your own team, gather a week of feedback and sharpen the instructions before you go broader. The first version is never the best version, and that's perfectly fine.

The three pitfalls we see most often

1. Starting too broad. An agent that answers "all HR questions" disappoints; an agent that knows the leave policies is useful straight away. Starting narrow and expanding works better than the other way round.
2. Connecting sources without checking permissions. The agent makes painfully visible what is shared too widely. Run an oversharing check on the connected sites first.
3. Not appointing an owner. An agent is not a project that is finished, but a product that needs maintenance: refreshing sources, refining instructions, processing feedback. Agree on who does that.

And when Agent Builder isn't enough?

Agent Builder is designed for question-and-answer agents on documents. If your agent needs to perform actions, talk to systems outside Microsoft 365 or run on your website, you'll end up with the full-fledged Copilot Studio. The knowledge you build up with Agent Builder — defining the purpose sharply, scoping the sources, testing with real questions — carries over there one to one.

Get started yourself — or together?

Building agents is a fixed part of our Copilot course: during the training, participants build their first agent on a real process from their own work. Curious which agents will deliver the fastest wins in your organisation? Schedule a no-obligation introductory call — we'll think along with you in 30 minutes.