Automation with Microsoft Copilot and Power Automate has become one of the fastest ways to streamline repetitive processes, connect applications, and reduce manual tasks without getting bogged down in code. If you're also working with Copilot Studio and agents, the next logical step is to create flows that these agents can execute in real time to provide much richer and more helpful responses to your users.
Throughout this guide, we'll explore how to create workflows with Microsoft Copilot, Copilot Studio, and Power Automate , the technical requirements you should consider, how to leverage natural language processing to design complex automations, and how to manage resource consumption. You'll see practical examples, such as a workflow that returns the weather forecast, Copilot prompt ideas, and best practices to ensure your workflows are fast, reliable, and easy to understand, even for non-technical users.
What are agent flows in Copilot Studio and why do they matter?
Agent flows in Copilot Studio are cloud-based flows designed to function as tools that an agent (your custom Copilot) executes within a conversation. Instead of simply generating text, the agent can call a flow, ask queries, execute business logic, and return the result to the user in a fully integrated way.
This type of flow consists of a specific agent trigger and at least one action , typically ending with a response to the agent itself. This allows the bot to orchestrate complex tasks and then use that information as part of the conversation with the user.
The great advantage is that these flows are deterministic and rule-based. With the same input, the result will always be the same, which is essential when we talk about automating critical steps in a process or applying sensitive business logic that cannot be left to the improvisation of a generative model.
In addition, Copilot Studio offers a unified environment where you can design, monitor, and analyze your agent workflows from a single dashboard. From there, you can see the name, description, status, performance, and even information to identify bottlenecks and opportunities for improvement.
Requirements for a flow to function correctly with an agent
For a cloud flow to function as an agent tool in Copilot Studio, simply creating it in Power Automate isn't enough. It must meet a series of very specific technical requirements to integrate correctly into the chat experience.
The first step is to use the appropriate trigger. The flow must begin with the "When an agent calls the flow" trigger . This type of trigger tells the system that the flow can be invoked by an agent, either from a specific topic or through generative orchestration.
Secondly, it is mandatory to include an output action of the type “Reply to agent” . This action sends the information back to the Copilot so that it can use it in the conversation, with parameters and variables that you explicitly define (for example, text, numbers, summaries, etc.).
Another key point is that the workflow must respond in real time, not asynchronously . In the "Respond to Agent" action settings, under "Network," the asynchronous response option must be disabled. Otherwise, the user will be left waiting indefinitely for a response that never arrives while the workflow continues running in the background.
Finally, time is also a factor. Flows must respond to the agent within a maximum of 100 seconds per action. To meet this limit, it's advisable to optimize internal logic, reduce unnecessary queries, and limit the amount of data returned to Copilot. Tasks that require more time can be placed after the "Respond to agent" action. This way, they will continue to run until the standard Power Automate limit (up to 30 days), but without blocking the conversation.
How to create an agent flow from Copilot Studio step by step
When you create a new agent flow directly from Copilot Studio, the system makes things much easier: the "When an agent calls the flow" trigger and the "Respond to the agent" action are already pre-configured in a basic template. This reduces errors and ensures compatibility from the very beginning.
The flow is linked to an agent topic in the form of a tool. Thus, when the user reaches a certain point in the dialogue or asks a specific question, the agent can decide to call that flow, pass it parameters, and use the result in its response.
Add a new empty agent flow to a theme
The first step in the process is to create the workflow skeleton and associate it with the topic that will trigger its execution. In short, the steps are as follows:
- Enter the page of Themes of the agent you want to work with.
- Open the topic from which you want to call the flow.
- Anywhere in the dialog tree, click on the icon of “Add node” and choose the option “Add a tool”.
- In the section of Basic tools, select “New agent flow”The agent flows page will open with a template ready to edit, which already includes the agent trigger and response action.
- Publish that flow immediately to save it, even if you haven't added your logic yet. You'll see a successful creation notification.
- Return to the agent using the corresponding button; a new node of type will appear in the topic AGENDA associated with the flow.
- Save the topic to confirm that the tool is linked to the agent.
This gives you what we could call a "shell" of agent flow ready for you to add the business logic you need: connector calls, conditions, loops, data processing, etc.
Configure actions and parameters in the flow (example with weather forecast)
Once the template is created, you can enrich the flow with specific actions. A very instructive example is that of an agent that returns the current weather forecast for a specific city using MSN Weather.
To do this, first return to the flow from the topic's action node using the "View flow details" link . On the overview screen, you can take the opportunity to rename it to something recognizable, such as "Get current weather forecast." This may seem like a minor detail, but a good name greatly simplifies maintenance when you have many flows.
In the flow designer, you edit the "When an agent calls the flow" trigger and define the input parameters that the agent can pass to it. For example, a text field for the city and a number field for the postal code . These parameters will then be reused in the internal actions.
Below the trigger, add a new action by finding the "Weather on MSN" connector and selecting "Get forecast for today." In the location field, insert dynamic content from the "City" parameter and do the same for the "Postal Code." This way, the action will query the forecast based on the user's query to the agent.
Finally, in the "Respond to agent" action, you define the outputs that the Copilot will use to construct its response. For example, you can add variables such as day_summary (text summarizing the day), location (text with the location), or chance_of_rain (number representing the probability of rain). Again, the goal is for the agent's response to be rich and structured, not just a generic message without context.
Once the inputs and outputs are configured, you publish the flow, and it's ready to be called from the theme or generative orchestration. The key is to clearly align input parameters, internal actions, and output fields so the agent can reason about the data consistently.
Management of connections and credentials in workflows linked to Copilot
Another aspect that shouldn't be overlooked is what connections and credentials flows use when running as part of an agent. In authenticated environments that allow it, flows can run using the credentials of the user interacting with the Copilot, which is especially useful when the information is sensitive or specific to each individual.
When using the flow in a supported channel, you typically don't need to configure anything special on the agent itself . Each user can manage their connections when using the bot, just as they would when running a flow independently in Power Automate.
However, in environments that use customer-managed keys (CMKs), there are still restrictions: cloud flows cannot be executed with customer credentials as part of generative orchestration or from themes. In these cases, it is important to verify that the "run-only" configuration of the flows is set to specific connections and not to the generic "User-provided run-only" option.
In practice, this involves coordinating with the platform administrator to align AI agent security , compliance, and user experience before exposing sensitive flows within a corporate Copilot.
Create cloud workflows with Copilot in Power Automate
Beyond agent flows, Copilot is integrated directly into Power Automate to help you design cloud flows using natural language . Instead of starting from a blank template and searching for connectors one by one, you can describe what you want to achieve and let Copilot generate an initial proposal.
The typical process starts by logging into Power Automate and making sure you're in the Home section of the sidebar menu. There you'll see a field like "Create your automation with Copilot" where you can write, in plain English, how your workflow should work.
The more detailed the description, the better. Copilot uses that information to suggest a trigger and one or more actions that match the request.
When you're happy with the proposal, you can keep it as is with the "Keep and Continue" option or request adjustments by adding more details in the text box. Copilot will regenerate the flow structure to adapt it to the new requirements. This way, the design becomes an iterative, conversational process rather than a traditional manual edit.
After accepting the suggested architecture, Power Automate will ask you to review and complete the necessary connections. You'll see which connectors have correct credentials (marked with a green symbol) and which require attention (with a warning icon). Once this is sorted, simply click "Create flow" and fine-tune the final details using the visual designer and the Copilot panel on the right.
How to write good requests for Copilot when creating flows
The quality of the workflow generated by Copilot depends largely on how well you formulate your request . It's not just about asking for something generic, but about clearly explaining what event triggers the workflow, what actions you want to be executed, and which connectors should be involved.
The "When X happens, do Y" pattern works very well . For example: "When an email arrives from [email protected] , save the attachment to a SharePoint library and post a message to the 'Sales' Teams channel with the file name." This type of structure greatly increases the chances of Copilot being successful on the first try.
It's also a good idea to be as specific as possible. Instead of saying "I want to process an email," you can specify that you want to send the email subject to a specific Teams channel or that you want to store certain data in a particular SharePoint list. The more precise you are, the less readjustment work you'll need to do later.
If you know which connectors you want to use, mention them directly in the request: Outlook, Teams, Forms, SharePoint, etc. Copilot can interpret those names and configure the appropriate actions for each one, saving you time and testing.
And if the first attempt isn't perfect, that's okay: the idea is to keep trying variations of the same request , fine-tuning details until you find the workflow you need. Copilot is designed precisely for that back-and-forth with the user.
Edit existing flows with Copilot's capabilities
Not only can you create new automations with Copilot, but you can also tweak existing flows in the new cloud flow designer . This experience is only available in the modern designer, not the classic one, so it's recommended that you migrate your flows if you want to take full advantage of it.
To use Copilot in an existing workflow, go to Power Automate, then to "My workflows," and select the edit option for the workflow you want to modify (either from the three dots or from the details page). The workflow will open with the Copilot panel on the right, ready to receive instructions in natural language.
From there, you can give it commands like “ delete action <action name> ”, “at the end of the flow I want to send an email with a subject equal to the SharePoint file name” or “instead of email, I want to post a message in a Teams channel”. Copilot translates these instructions into actual changes in the flow diagram.
It's also possible to set up more advanced scenarios. For example: "Check if each returned SharePoint item has a title 'USB' and, if so, send an email." This way, you can introduce conditions and filtering logic without having to build them piece by piece in the visual interface. Although you can always review and adjust the configuration manually afterward.
This combination of conversation-guided editing and detailed visual control makes maintaining complex workflows much more manageable, even for business teams that don't consider themselves technical.
Hyperautomation, AI and low-code platforms: the context of Copilot
Behind all these capabilities lies a broader trend: hyperautomation , which involves automating as much as possible within an organization by combining AI, RPA, low-code platforms, and process mining. Microsoft Copilot and Copilot Studio are situated precisely at this intersection.
The artificial intelligence integrated into these tools allows automation to go beyond simply following a fixed script . Instead, it learns from data, recognizes patterns, adapts to changes, and makes decisions in real time. Agents powered by cloud workflows can handle increasingly complex tasks without requiring constant manual intervention.
Low-code or no-code platforms make it easier for non-technical users, such as business analysts or operators, to design and manage their own workflows. Thanks to drag-and-drop interfaces, templates, and AI wizards, the barrier to entry is lowered, and the IT team is no longer the bottleneck for every small change in a process.
Furthermore, process and task mining offers an objective view of how processes are actually executed in practice. With this data, it's easier to see where it's worthwhile to invest in automation, measure the impact of changes , and continue iterating to improve results.
All of this points to the idea that Copilot workflows aren't just small, isolated utilities, but pieces of a larger machine that helps organizations become more agile and scalable. Perfectly prepared for whatever comes next.
Agent flow capacity usage and licensing in Copilot Studio
When designing agent workflows, it's important to keep in mind that each action performed consumes Copilot Studio capacity . This is a significant detail, because as your automations grow, so will the associated resource consumption, and it's crucial to plan for this.
From the Power Platform admin center, under Licensing > Copilot Studio , you can review how many agent flow actions each flow has consumed. This gives you a clear view of which processes are the most capacity-intensive and where you could optimize them.
Interestingly, the tests you run from the designer don't consume Copilot Studio's capacity. However, they can consume resources from other billable services, such as specific generative models. This allows you to iterate on your design with relative peace of mind before deploying the workflow to production.
The cost also depends on how the flow is invoked. If it's run from an agent theme, it consumes a standard agent response plus the agent flow actions. If it's called using generative orchestration, it counts as a standalone action plus the flow actions. And if you test it from the test chat embedded in the agent, only the agent flow actions are counted, not the direct chat messages.
Understanding these rules allows you to accurately determine the necessary capacity and avoid surprises , especially in projects where workflows are executed frequently or are exposed to a large number of users.
Convert Power Automate flows into agent flows
If you already have automations in Power Automate that work well, you might not want to rebuild them from scratch. The good news is that you can convert an existing cloud flow into an agent flow , so it's managed in Copilot Studio and uses its capacity instead of Power Automate's.
To do this, you must ensure that the flow is part of a solution. On the flow details page in Power Automate, if the "Solutions" tile appears , it means it's already included. If not, you'll need to add it to a solution before proceeding.
Another requirement is that the workflow must be in the Power Platform environment where you want to use Copilot Studio capabilities. If this capability is not yet assigned, you will need to coordinate with your administrator to configure it correctly.
Once those conditions are met, you edit the workflow and change its plan to “Copilot Studio” . After saving, the system will ask you to confirm that you want to convert it; this is an important step because it is a one-way operation that cannot be undone , precisely because of the associated change in billing model.
After the conversion, you can use that flow as part of your agents, take advantage of Copilot Studio's tracking and capabilities, and continue to access it from Power Automate as well, offering considerable flexibility for hybrid scenarios.
Activate or deactivate Copilot in Power Automate depending on the region
Copilot availability in Power Automate can vary depending on your tenant's region and the availability of GPUs in that region. In locations with available GPU resources (such as the UK, Australia, the US, or India), Copilot is typically enabled by default at the tenant level. To disable it, an administrator must contact support and use a PowerShell script to disable Copilot globally for the tenant, as there is currently no fine-grained control per environment.
In regions without GPUs (except in sovereign clouds), Copilot is also enabled by default, but through the cross-region data sharing mechanism . If you need to disable it, you can do so by disabling this data sharing from the Power Platform admin center, also at the tenant level.
Whatever your situation, it's advisable to align this configuration with your organization's compliance, privacy, and governance policies and Zero Trust principles in the age of AI before massively deploying cloud flows with Copilot and agents in production.
The combination of agent flows, Copilot in Power Automate, Process Mining, and integration with low-code platforms makes Microsoft Copilot a powerful tool for transforming processes: it allows you to move from isolated automations to a coherent ecosystem of agents, flows, and analytics that anyone in the company can understand, adjust, and scale with relative ease, provided that technical requirements, capacity management, and good request design and security practices are taken into account.
