Architect Foundations · 18% of the exam

tool design and mcp integration

Domain 1 covered how an agent decides its next step. Domain 2 covers the tools it chooses from: how each one is described, what it returns when it fails, which agent holds it, and where its server is configured.

The scenario this domain runs on

You are building a customer support resolution agent with the Claude Agent SDK. It handles high-ambiguity requests such as returns, billing disputes and account problems. It may refund up to $500 on its own, and it reaches your backend only through four custom tools on your team's support MCP server:

get_customerlookup_orderprocess_refundescalate_to_human
80%+first-contact resolution. A delivery complaint sent to get_customer, a timeout reported as "Operation failed", or a refund answered with a discount code each sends the customer back a second time.

Each section below is one way the tool layer lets this agent down. In 2.1 vague descriptions send a request to the wrong tool. In 2.2 a failed tool says nothing the agent can act on. In 2.3 the agent holds tools outside its role and misuses them. In 2.4 the team's MCP servers sit in a file teammates never get. In 2.5 an agent working on a codebase picks the wrong built-in tool.

Scenario switch2.5 runs on a second scenario, the guide's Developer Productivity with Claude. Built-in tools act on a codebase, so 2.5 follows an agent in a legacy billing system. 2.1 and 2.3 also walk the guide's research-system examples, each in a short flagged note.

Start hereThe words used throughout this domain

MCP
Model Context Protocol, an open standard through which servers offer an AI application tools and resources.
MCP server
The process that owns a real capability, such as your order database or Jira, and offers tools for it: named actions like lookup_order, which Claude chooses and your application calls. Your team writes the support server, while Jira usually comes from an existing community server.
tool_use, tool_result
The Claude API block that requests a tool call, and the block that returns its outcome. is_error is the result's failure flag. MCP spells it isError, the form 2.2 uses.
subagent, coordinator
A subagent is a separate Claude session with its own context window, set up by an AgentDefinition, the SDK object listing its prompt and tools. The coordinator delegates to it.
hook
Code that runs before or after a tool call and can block it. Domain 1 covers hooks.
Grep, Glob
Built-in Claude Code tools: Grep searches file contents, and Glob matches file paths. 2.5 covers both.
context cost
Tokens loaded before your first message, such as tool definitions. MCP tool search in Claude Code defers most of them (2.4).

task statements

flashcards

0 / 15 known
Card 1 of 15