Table of Contents
Copilot Studio is Microsoft’s tool to create our own agents. Copilot Studio agents can respond to questions based on business data and can also perform actions like sending emails, adding a row to a table, or querying data if we add tools and instructions to them.
All the above works well in some cases where the agent can use “common sense” and context to perform actions, but agents can also be unpredictable for critical business processes. For some use cases require a more deterministic approach: we need an agent to follow a specific set of steps and a specific set of rules or conditions: here is when Copilot Studio agent flows come in handy!
To find Agent flows in Copilot Studio, simply open the Copilot Studio portal and on the left panel you will see a “Flows” section: 
Check out our agentic automation blog to learn more!
Introducing Agent Flows in Copilot Studio
The easiest way to describe Agent Flows is “Power Automate workflows inside Copilot Studio”. When we work with Power Automate, we can create a set of steps to follow, so by bringing Power Automate capabilities to Copilot Studio we can enhance agents with:
- Deterministic workflows
- Pre-defined structure of steps, rules or conditions
- Repeatable tasks performed with consistency
The interface to build Agent Flows looks very similar to Power Automate, and it has almost the same set of actions and connectors. 
From Copilot Studio’s agent flows, there are a few more capabilities that can be leveraged:
- Human in the loop advanced approvals
- Python code interpreter
Key Components of Agent Flows Architecture
Understanding how Agent Flows work starts with knowing what they’re made of, and as we have mentioned above since Agent flows are similar to Power Automate workflows, the components these two are pretty similar (to learn more about the differences between these tools, read our blog on this topic).
Triggers
A trigger is something that needs to happen for an agent flow to start:
- Receive an email
- Add a row in a Dataverse table
- The first day of the month (scheduled)
- A user calls the flow manually
- An agent calls the flow
Actions
Once the workflow starts, it needs to perform a sequence of actions. The number of actions we add to an agent flow depends on the kind of logic we want to perform, but the list of things an agent flow can vary within a large variety of actions, from simple to complex:
- Add a row to SharePoint list
- Update a row in Dataverse when an order is created
- Send a Teams notification
- Send an approval request
- Create a new customer in a third party system (HubSpot, Salesforce)
- Post an invoice in your ERP system
Agents vs Agent Flows
Since agents leverage LLMs, an Agent can have conversations and interpret natural language based on contexts, so it can try to choose an appropriate next step based on what it knows. On the other hand, Agent flows can perform a set of actions that configure step by step, so it has a more deterministic behevious.
With agents, the same input could potentially lead to different outputs. With Agent flows, since we’re building consistency and predictability, the same input will lead to the same output.
| Agent | Agent Flow |
Purpose | Conversational reasoning and context | Business process automation |
Behavior | Dynamic, adaptive | Deterministic, rule-based |
Runs | During a conversation or autonomously if a trigger is connected to the agent | When triggered by an event, a user or agent, or by schedule |
Licensing | Copilot Studio capacity | Copilot Studio capacity (per action) |
When combined, Agent Flows and Agents can offer a good mix of agility, adaptability and predictability.
How Agent Flows enable orchestration
Reusability across multiple agents
Agent Flows aren’t tied to a single agent, they can be reused across multiple agents. This ensures consistency and simplifies process management. For example, if we build an agent flow that processes expense reports can be connected into other agents such as finance agent, an HR agent, and a manager approval agent without rebuilding it each time.
Deterministic reliability within a context-based, dynamic environment
Agents are flexible, context-based, and adaptive by nature, which is good for some scenarios but can become unpredictable for critical processes. Agent Flows are deterministic because they execute the actions we set up in the workflow, following a rule-based path, so the same input always produces the same output, making them reliable and predictable.
Human-in-the-Loop orchestration
For decisions that require human judgment and/or manual review, Agent Flows support multistage approvals where different stages can be set up in a single action and let the automation decide on each stage based on the output from the previous stage. 
Are you ready to discover the joy of automation?
If your organization is already using Copilot Studio or Power Automate, Agent Flows should also be on your roadmap, and you don’t need to figure this out on your own. Contact us to learn more about how our Copilot Studio and Agentic automation consulting services can help you and your team get work done!
