Tool Design, MCP, and Agent Context

What this chapter covers

Tool Design, MCP, and Agent Context is 13% of the Databricks Context Engineer Associate exam. It covers how tools, their descriptions and their outputs consume the context window, and how to keep that cost low without hurting the agent's choices.

The chapter has two parts: progressive disclosure (staged tool discovery over MCP, and packaging rarely used capabilities as Agent Skills), and working with individual tools (picking the right Unity Catalog tool, fixing descriptions the agent confuses, and clearing raw tool outputs when the window fills).

Questions are single-answer and multiple-selection ("Which TWO ...?") multiple choice. Almost all are scenarios that describe an agent, a trace, a configuration or MLflow evaluation results and ask for the diagnosis or the remediation that meets a stated constraint. There are no exhibits.

How the topics fit together

Tools enter the window in three ways: their descriptions, their schemas and their outputs. Each stage of the chain lets less of that in, and each stage has its own failure.

• Disclosure questions ask how to cut token use when an MCP server exposes many tools.
• Skills questions ask which system prompt instructions to move into Agent Skills and how they load.
• Selection questions ask which Unity Catalog tool fits a task by function, inputs and outputs.
• Ambiguity questions ask which overlap in two descriptions makes the agent pick the wrong tool.
• Clearing questions ask which old tool outputs can be removed from a deep history.

How to use these pages

Each topic page has the same shape, so you can skim for what you need.

• What it is: the idea in plain words.
• How it works: the mechanism, with tables and examples.
• What to remember for the exam: the facts to memorize and how questions are phrased.
• Common mistakes: the classic traps.
• Check yourself: three questions to test your understanding.