| name | foundry-tool-governance |
| description | This skill should be used when designing Microsoft Foundry tool surfaces: Foundry Toolbox, tool catalog configuration, MCP servers, Azure Functions, OpenAPI tools, API Management AI Gateway policy, tool versioning, and allowed-tool scoping. |
Foundry Tool Governance
Tool governance is the Microsoft-native implementation of the core tool-design principle: keep the agent's callable surface small, unambiguous, permissioned, observable, and versioned. Foundry Toolbox is the preferred abstraction when a team needs a curated tool set rather than a loose pile of MCP servers and OpenAPI definitions.
When to Activate
Activate this skill when:
- Packaging several tools into a Foundry Toolbox.
- Choosing between built-in Foundry tools, Azure Functions, OpenAPI, MCP, or API Management AI Gateway.
- Scoping which MCP tools an agent may call.
- Versioning and promoting a tool catalog.
- Adding policy, authentication, logging, or throttling around agent tools.
Do not activate this skill for adjacent work owned by other skills:
- Writing generic tool descriptions, schemas, or error payloads: core
tool-design.
- Designing identity flows for tool calls:
azure-identity-for-agents.
- Designing retrieval quality or knowledge-source chunking:
azure-agentic-retrieval.
- Designing human approval boundaries around autonomous action: core
harness-engineering.
Core Concepts
Foundry Toolbox should be treated as the curated boundary between the non-deterministic agent and deterministic systems. The toolbox is not just a convenience wrapper. It is where tool selection, versioning, authentication, and governance become explicit.
Use this hierarchy:
| Need | Microsoft-native default |
|---|
| Built-in knowledge or action | Foundry tool catalog |
| Curated cross-agent tool set | Foundry Toolbox |
| Custom synchronous action | Azure Functions MCP or OpenAPI |
| Existing REST estate | API Management AI Gateway or OpenAPI tool |
| Central policy and throttling | API Management AI Gateway |
| Long-running action | Queue-backed function or workflow with status polling |
Practical Guidance
Design a toolbox in this order:
- List the workflows the agent must perform, not the APIs available.
- Remove overlapping tools before exposing them to the model.
- Group tools by stable domain boundaries such as
crm, devops, search, or billing.
- Put authentication and authorization in Foundry connections, managed identity, OBO, or APIM policy.
- Give every tool an actionable error response that tells the agent whether to retry, ask, or stop.
- Create a new Toolbox version for meaningful changes, test it, then promote it.
- Review traces to confirm the agent calls the intended tool for representative tasks.
Examples
Example: toolbox boundary
Avoid: agent sees twenty raw CRM endpoints.
Prefer: toolbox exposes search_customer, summarize_customer_context, create_follow_up_task.
Example: APIM governance
Agent -> Toolbox MCP endpoint -> API Management policy -> internal API
The gateway can enforce token limits, route versions, redact headers, and centralize logs without changing the agent prompt.
Guidelines
- Prefer Foundry built-in tools before custom tools when the capability matches.
- Prefer Toolbox when multiple agents should share the same curated tool set.
- Prefer APIM AI Gateway when policy, throttling, or existing REST API exposure matters.
- Expose workflow-level tools, not raw backend implementation details.
- Version tool catalogs deliberately and promote only after trace-based validation.
Gotchas
- Toolbox does not fix bad tool descriptions: A curated endpoint still fails if descriptions overlap or omit recovery guidance.
- Raw MCP expansion creates routing noise: Importing every MCP tool is convenient but can degrade selection. Scope tools before exposing them.
- Policy outside the prompt still shapes behavior: Rate limits, auth failures, and blocked calls must return agent-actionable errors or the model will retry blindly.
Integration
- core tool-design - Owns the generic tool contract and consolidation principle.
- azure-identity-for-agents - Owns managed identity, OBO, and RBAC choices.
- responsible-ai-guardrails - Adds safety checks and prompt-injection controls around tool use.
- core evaluation - Measures tool call accuracy and task completion.
References
Skill Metadata
Created: 2026-05-24
Last Updated: 2026-05-24
Author: Agent Skills for Context Engineering Contributors
Version: azure-0.1.0