Tools your agent can actually trust
Microsoft Developer walks through a practical pattern for making agent tool use safer and more governable using Azure AI Foundry’s Toolbox.
Overview
The video uses a “Sparkle Cupcakes” agent scenario to show how an agent can be given a governed set of tools:
- The agent can read information (like menu and order status) freely.
- The agent can act (like placing orders or issuing refunds) only when those actions are gated behind approval/confirmation.
It also covers how to centralize and reuse tool definitions across multiple agents, and how enabling tool search can help keep token costs flat as the toolbox grows.
Key segments
Managing AI agent components (0:00)
Introduces the idea of breaking agent capabilities into components, with special focus on tool access and governance.
What is Foundry Toolbox? (1:05)
Defines Foundry Toolbox as a governed place to manage the tools an agent is allowed to use.
Create a centralized toolbox (1:39)
Shows creating a single, centralized toolbox so multiple agents in the same “shop” can reuse the same tool definitions.
Add MCP and web tools (2:12)
Demonstrates adding:
- MCP tools (Model Context Protocol)
- Web tools
Configure guardrails and skills (2:46)
Covers configuring controls so that:
- Read-only operations remain available without friction.
- High-impact operations (like refunds) require explicit approval.
Publish the toolbox (3:20)
Publishes the toolbox so it can be consumed consistently by agents.
Connect the toolbox to code (3:51)
Connects the published toolbox into an application/code workflow so the agent can use the governed tools.
Centralized governance and management (4:25)
Wraps up with the governance model:
- Define the toolbox once.
- Reuse it across agents.
- Separate “reading” from “acting.”
- Ensure nothing happens without confirmation.