Compare agent tools
An agent powered by the GitHub Copilot harness supports a range of options for tools that enable you to integrate with enterprise systems. This unit provides an overview of each type and when to consider using it within your solution.
Use a connector for a supported service
Power Platform connectors expose actions from hundreds of Microsoft and third-party services, such as SharePoint, Outlook, and Dataverse. Each action, such as retrieving a record or creating an item, lets the agent read from or write to the service. Because the connector already exists, you configure it instead of building an integration.
A connector is usually the first option to consider when:
- The system is a supported service. Check that the connector includes the specific action the need calls for. For example, a connector for a service management system might support retrieving records but not the update your team requires.
- The agent should work with only one defined resource and operation. A specific connector action keeps the agent's reach narrow, such as retrieving items from one approved SharePoint list. A broad option such as Work IQ draws on the wider Microsoft 365 data that a user has permission to see.
When a connector offers several actions, you choose which ones the agent can use.
Use an MCP server for custom or internal services
Model Context Protocol (MCP) is a standard way to expose tools to an agent. An MCP server can expose capabilities from custom services, databases, or internal APIs. MCP servers on this harness are currently in preview.
An MCP server fits when:
- No suitable connector exists for the system.
- A team already runs an MCP server for the system. For example, if the utility's IT team runs an MCP server in front of the internal case system, your agent can use the server's case-lookup tool.
- Several agents share the same tools, or the underlying APIs change often. The server's owner maintains the tools in one place and changes a tool once on the server instead of in each agent.
If no server exists yet and a connector already covers the need, the connector is usually quicker to set up.
Before you recommend an MCP server, keep these points in mind:
- The server's tools come with it. When you add a server to an agent, every tool the server exposes becomes available to the agent. You can turn off individual tools the agent doesn't need.
- The server's author writes the tool descriptions. The quality of each tool's name and description affects how reliably the agent chooses it.
- Only a limited number of servers can run at once. An agent can connect to more than one MCP server, but the number that can run at the same time in a conversation is capped. If the agent tries to use more servers than that in one turn, the extra servers are skipped for that turn, so attach only the servers the agent needs.
Use a workflow for a fixed, multistep process
Workflows are automated flows that you build in the Copilot Studio flow designer. When the agent calls a workflow, the workflow runs a defined sequence of steps, such as an approval, a data transformation, or other business logic. Workflows suit repeated, deterministic processes that must run the same way every time.
For example, suppose submitting an approved case update also means notifying the requester and assigning the case to a queue, in that order, every time. A workflow keeps those steps consistent. The agent decides when to run the workflow, and the workflow defines what happens inside it. As a rule, leave ambiguous requests to the agent's reasoning, and put steps that never change into a workflow.
Use an IQ tool to bring in context your organization already manages
Foundry IQ, Fabric IQ, and Work IQ mainly bring context into the conversation. Each one connects the agent to a Microsoft platform where your organization already manages the data, so the agent can answer questions or produce insights from it. You add each one to an agent as a tool. The names are similar, but each fits a different kind of need:
| Tool | Connects the agent to | Fits needs such as |
|---|---|---|
| Foundry IQ | A knowledge base that a team built and tuned in Microsoft Foundry | Questions over curated content that another team maintains and continues to tune |
| Fabric IQ (preview) | Business data and analytics in Microsoft Fabric, grounded in your organization's semantic model | Analytical questions about business data |
| Work IQ (preview) | Microsoft 365 context, such as email, calendar, and Microsoft Teams, through MCP tools | Requests that depend on a person's Microsoft 365 work context |
Work IQ is read-only unless an administrator turns on write operations in the Microsoft 365 admin center.
For example, if the utility's field operations team already maintains a tuned knowledge base of repair procedures in Foundry IQ, your agent can use it for procedure questions instead of you adding the same documents as a knowledge source.
Decide which type of tool fits
For each need that calls for a tool, ask these questions:
- Is the system a supported service with a connector that includes the needed action? Start with the connector.
- Is the system custom or internal, with an MCP server that provides the needed tools? Add the MCP server as a tool.
- Does the need start a fixed sequence of steps? Use a workflow.
- Does the need depend on a tuned knowledge base in Foundry, business data in Fabric, or a person's Microsoft 365 work context? Consider Foundry IQ, Fabric IQ, or Work IQ, respectively.
Applied to the utility scenario's case management agent, the reasoning looks like this:
| Request | Reasoning | Recommendation |
|---|---|---|
| Explain the escalation policy | Reference content that your team maintains | Knowledge source with the approved policy documents |
| Look up a case's current status | Live read from an internal system that an MCP server exposes | The server's case-lookup tool |
| Submit an approved case update | A write that also notifies the requester and assigns the case, in a fixed order | Workflow that uses a supported update action |
| Assess an unusual exception | Specialist judgment that another team's agent provides | Connected agent |
Each recommendation is a starting point. Before it's final, you confirm that the agent can connect to each option and is permitted to do what the request needs.
Reflect: Think of one task that an agent you work on needs a tool for. Which type of tool fits it best, and what would you check before you recommend it?