Understanding Microsoft Copilot Studio: Agents, Workflows, and Harnesses
Introduction
Microsoft Copilot Studio is a graphical, low-code platform for building and managing AI-powered agents and workflows.
It provides a unified environment where organizations can create agents, automate business processes, connect organizational data and systems, and publish solutions to the channels where users already work.
One of the most important ideas in the current Copilot Studio architecture is that the platform is no longer only about creating conversational bots.
Copilot Studio brings together several building blocks:
- Agents
- Workflows
- Agent flows
- Knowledge
- Instructions
- Tools
- Connectors
- Harnesses
- Analytics
- Evaluations
- Administration
These components can be used independently or combined to create more sophisticated business solutions.
What Is Microsoft Copilot Studio?
Microsoft Copilot Studio is designed to provide a single studio for building and managing AI agents and workflows.
Because the platform is low-code, solutions can be created without requiring extensive software development experience.
At the same time, professional developers and solution architects can build more sophisticated integrations using organizational data, connectors, tools, and enterprise systems.
A useful conceptual model is:
Copilot Studio
↓
Agents + Workflows
↓
Knowledge + Instructions + Tools + Logic
↓
Enterprise Data and Systems
↓
Business Solutions
The important point is that Copilot Studio is the platform where these different capabilities can be designed and coordinated.
Agents
An Agent is an AI assistant capable of handling conversations and completing tasks.
According to Microsoft, an Agent can use three particularly important elements:
Instructions + Knowledge + Tools
Instructions define how the Agent should behave.
Knowledge Sources provide information that the Agent can use.
Tools allow the Agent to take action.
The Agent can reason about a request and determine an appropriate next step based on its instructions and available context.
Agents can interact with employees or customers through several channels, including Microsoft Teams, Microsoft 365 Copilot, websites, mobile applications, and other experiences.
This gives us a useful architectural model:
User Request
↓
Agent
↓
Instructions
↓
Knowledge / Context
↓
Reasoning
↓
Tool, when necessary
↓
Response or Action
Agents can be created using natural language descriptions and tested before publication.
Some Agents can also operate proactively as participants in business processes, such as employee onboarding or recurring business activities.
Workflows
Copilot Studio also supports Workflows.
A Workflow is an automation built using a drag-and-drop designer.
Users can describe the automation they want, and Copilot Studio can help generate, configure, and validate the workflow.
An important characteristic of Workflows is the combination of traditional automation with AI capabilities.
A workflow can combine:
Agents + Tools + Logic
This provides the predictability normally associated with business process automation while introducing the flexibility of AI where appropriate.
Workflows also support capabilities such as testing and human-in-the-loop controls.
This distinction is important because not every business process should depend entirely on AI reasoning.
Some operations require predictable and controlled execution.
Agent Flows
Copilot Studio also includes Agent flows.
Agent flows are the established flow format in Copilot Studio and provide an authoring experience similar to Power Automate.
An Agent flow can operate as a standalone automation or be attached to an Agent as a Tool.
When attached to an Agent, the flow can perform an operation and return its result to the Agent.
Conceptually:
User
↓
Agent
↓
Tool
↓
Agent Flow
↓
Business System
↓
Result
↓
Agent
↓
User
Agent flows can also run prompts, call Agents, and include human review steps.
For Microsoft 365 scenarios, this creates interesting possibilities for integrating Agents with enterprise processes.
For example, an Agent could eventually use a flow to interact with SharePoint or another connected business system.
The Concept of a Harness
One of the most important concepts in the current Copilot Studio architecture is the Harness.
Microsoft describes the Harness as the engine that carries out the work behind the scenes.
The selected Harness affects several characteristics of the solution, including:
- how it reasons;
- the complexity of tasks it can perform;
- its built-in capabilities;
- billing.
Copilot Studio currently identifies three Harness options:
GitHub Copilot harness
Designed for reasoning-heavy, multi-step work and complex business processes.
Standard harness
Designed for rule-based Agents and structured, repeatable conversations.
Copilot chat harness
Designed for extending Microsoft 365 Copilot Chat with organizational knowledge.
This distinction is extremely important when studying modern Copilot Studio because two Agents created for different Harnesses can behave differently even though both are created within Copilot Studio.
GitHub Copilot Harness
With the GitHub Copilot harness, the solution reasons from the Instructions and Knowledge provided to it.
It interprets the request, selects appropriate Tools, and determines the steps required to complete the task.
The developer therefore does not necessarily need to explicitly script every possible conversational path.
A simplified model is:
Request
↓
Instructions + Knowledge
↓
Reasoning
↓
Tool selection
↓
Multi-step execution
↓
Result
This approach is particularly relevant to tasks that require more autonomous reasoning and multiple steps.
Standard Harness
The Standard harness follows a different model.
For Agents using the Standard harness, natural language understanding can match a user’s request with a Topic.
A Topic represents a portion of a conversation designed with connected steps, questions, conditions, and other logic.
Conceptually:
User Request
↓
Natural Language Understanding
↓
Topic
↓
Defined conversational logic
However, Topics are not the only mechanism available.
When a request falls outside the defined Topics, the Agent can still generate conversational responses using its connected Knowledge Sources.
This makes the Standard harness particularly interesting because it can combine structured conversational behavior with generative answers.
Copilot Chat Harness
The Copilot chat harness focuses on connecting organizational knowledge to Microsoft 365 Copilot Chat.
Employees can receive answers grounded in organizational content while remaining inside their normal Microsoft 365 Copilot experience.
Conceptually:
Microsoft 365 Copilot Chat
↓
Agent
↓
Enterprise Knowledge
↓
Grounded Answer
This is particularly relevant when organizations want to specialize Microsoft 365 Copilot around a particular domain or collection of enterprise knowledge.
Building with Natural Language
Copilot Studio also introduces a natural-language approach to solution creation.
Instead of always starting from an empty designer, a maker can describe a business objective.
Copilot Studio can then determine an appropriate combination of Agents and Workflows.
The solution can subsequently be refined through additional conversation and edits.
This capability is currently documented by Microsoft as preview.
How Agents Handle Requests
The way an Agent processes a request depends significantly on its Harness.
This means that the following question is becoming increasingly important when designing Copilot Studio solutions:
Which Harness is appropriate for this problem?
For example, a highly structured process with predictable conversational paths might fit the Standard harness.
A complex multi-step task requiring greater reasoning could require the GitHub Copilot harness.
A scenario focused on extending Microsoft 365 Copilot Chat with organizational knowledge could fit the Copilot chat harness.
Therefore, the Harness should be considered part of the architecture rather than merely an implementation detail.
Managing Production Solutions
Creating an Agent is only one part of the lifecycle.
Copilot Studio also provides capabilities for operating solutions after they have been created.
Microsoft highlights three important areas.
Analytics
Analytics can be used to monitor performance, custom metrics, and session outcomes.
Evaluations
Evaluations can help validate Agent quality using test sets and graders before and after publication.
Administration
Administration capabilities help organizations inventory, secure, and manage Agents, including role-based access and cost management.
This is particularly important for enterprise environments.
A production Agent requires more than good Instructions and Knowledge.
It also requires:
Testing + Evaluation + Security + Administration + Monitoring + Cost Management
Extending Microsoft 365 Copilot
Copilot Studio can also be used to extend Microsoft 365 Copilot.
Organizations can add their own:
- Instructions;
- Tools;
- Knowledge.
This makes it possible to customize Microsoft 365 Copilot for particular tasks or areas of organizational knowledge.
Copilot Studio can therefore participate both in the creation of specialized Agents and in the extension of the broader Microsoft 365 Copilot experience.
A Useful Architecture Model
Based on Microsoft’s current Copilot Studio model, we can build the following mental map:
Microsoft Copilot Studio
↓
Agent / Workflow
↓
Harness
↓
Instructions + Knowledge + Tools + Logic
↓
Reasoning / Orchestration
↓
Enterprise Systems
↓
Business Outcome
The Harness influences how this process is executed.
This is an important evolution from thinking about Copilot Studio simply as a chatbot development environment.
Applying This to SharePoint
For our SharePoint-focused learning path, these concepts provide an excellent foundation.
A simple knowledge Agent might eventually follow this architecture:
User
↓
Copilot Studio Agent
↓
Instructions
↓
SharePoint Knowledge
↓
Grounded Response
Later, we can introduce Actions:
User
↓
Agent
↓
Tool
↓
Agent Flow
↓
SharePoint
↓
Result returned to the Agent
This demonstrates an important distinction that we will maintain throughout our learning journey:
Knowledge helps the Agent know something.
Tools help the Agent do something.
What We Should Learn First
The Copilot Studio overview introduces many concepts, but they should not all be learned simultaneously.
A practical learning progression is:
Agent
→ Instructions
→ Knowledge
→ Grounding and Retrieval
→ Testing
→ Tools
→ Agent Flows
→ Workflows
→ Harness selection
→ Security and Governance
→ Enterprise Architecture
For our SharePoint-focused journey, the first major objective should be understanding how an Agent uses Instructions and SharePoint Knowledge before adding Actions.
Conclusion
Microsoft Copilot Studio is a low-code environment for building and managing AI-powered Agents and Workflows.
Its current architecture extends considerably beyond traditional conversational bots.
Agents can use Instructions, organizational Knowledge, and Tools to reason about requests and complete tasks.
Workflows provide automation capabilities that combine logic, Agents, and Tools.
Agent flows provide another mechanism for implementing automations and exposing them to Agents as Tools.
Finally, Harnesses define important aspects of how these solutions reason and execute.
The three Harnesses currently described by Microsoft are:
GitHub Copilot harness — reasoning-heavy and multi-step scenarios.
Standard harness — structured and repeatable conversational scenarios.
Copilot chat harness — extending Microsoft 365 Copilot Chat with organizational knowledge.
Understanding these concepts gives us the foundation required before moving into deeper topics such as Knowledge Sources, Grounding, Retrieval, Tools, Actions, Agent Flows, SharePoint integration, security, governance, and enterprise deployment.
Official Microsoft Learn Reference
Microsoft — Copilot Studio Overview
edvaldo b. guimarães filho 