Implement generative AI and agentic solutions 2.2 Build agents by using Foundry
A toolbox has grown to about 300 tools.
Study guide skill: Integrate agent tools, including APIs, knowledge stores, search, content understanding, and custom functions
Question B2-010
A toolbox has grown to about 300 tools. Every model request now carries all the tool definitions, input-token costs are rising, and the agent increasingly picks a similar but wrong tool. Which toolbox capability (currently in preview) addresses this?
Select one answer.
Answer: A
Tool search (preview) hides tools by default and exposes only two meta-tools: tool_search, which returns the most relevant tools for a described need, and call_tool, which invokes a discovered tool by name. You can pin critical tools so they're always available. A bigger context window still sends every definition, and temperature or approval settings don't reduce token cost or improve selection.
Deep dive
How it works
A big toolbox runs into three problems that the toolbox overview spells out. Every tool definition adds input tokens, whether the model uses that tool or not. Tool definitions compete with conversation history and other context for space in the context window. And when a model must choose from hundreds of tools, it's more likely to select a similar but incorrect tool.
Tool search (preview) fixes this by changing what the model sees. It hides tools by default and exposes only two meta-tools:
- tool_search: describe what you need and get back the most relevant tools.
- call_tool: invoke any discovered tool by name.
You stay in control of what surfaces. You can pin critical tools so they're always available, add context so discovery uses your organization's terms, and auto-pin frequently used tools. The overview's summary: a single toolbox can hold hundreds of tools without flooding the model's context, inflating token cost, or degrading selection accuracy.
The other options, and when you'd use them
- A model with a larger context window: The cost problem stays. Every tool definition still adds input tokens on every request, and selection among hundreds of tools stays error-prone.
- require_approval set to always on every tool: Approval is a safety control. The MCP guidance is to require approval for high-risk operations. Approving calls doesn't shrink the tool list or help the model pick the right tool.
- Raising the temperature: Temperature changes output randomness. A higher value makes output more random and divergent, which is the opposite of more accurate tool selection, and it doesn't touch token cost.
Exam tip
"Hundreds of tools, rising token cost, wrong tool picked" = tool search: tools hidden by default, tool_search + call_tool, pin the critical ones.
References
Last reviewed: