Tools overview for agents

Tools extend what your agent can do beyond answering questions from knowledge sources. They let your agent call APIs, run automated flows, interact with external services, and retrieve real-time data.

What are tools?

Tools are external capabilities your agent can invoke during a conversation.

When a user asks the agent to do something that requires an action, the agent orchestrator selects a tool based on the conversation context and the tool's description.

Why add tools to your agent?

Add tools to give your agent access to external services and workflows.

With tools, your agent can:

  • Look up real-time data from external systems, such as Customer Relationship Management (CRM), IT Service Management (ITSM), and databases
  • Take actions on your behalf, such as creating records, sending notifications, and updating data
  • Integrate with third-party services and APIs
  • Run automated workflows

For built-in file and data capabilities, learn more in Files the agent creates and Built-in skills.

How the agent decides to use a tool

The agent orchestrator evaluates each user message and decides whether a tool is needed.

The tool's name and description are critical. Well-described tools are invoked more accurately.

Your agent's instructions also influence when tools are used.

You don't need to create explicit triggers or topic flows for each tool call - the orchestrator handles this automatically based on the conversation.

Limits

A few limits apply to tools and to files that tools produce:

  • Files the agent creates during a turn are capped at 10 MB per file. Files that exceed the cap aren't returned to the user. Ask the agent to split large output into smaller files.
  • Model Context Protocol (MCP) servers count against the total number of tools an agent can host. The number of MCP server instances that can run concurrently in one conversation is also capped, so keep the count of attached servers small and remove servers you no longer use. See Add an MCP server (GitHub Copilot preview) or Add an MCP server (Customer Conversation).

These limits apply to agents powered by both the GitHub Copilot harness and the Customer Conversation harness.

  • Last updated on