Learn · Tool design & MCP · CCAR-F task 2.3

Distributing tools and tool_choice

Give each agent only the tools its role needs, constrain generic tools, and use tool_choice when a step must happen.

Key ideas

• Too many tools degrade selection: the guide contrasts 18 tools with 4-5.
• Scope tools per role; agents misuse tools that are outside their job.
• Give a narrowly scoped tool (such as read_source) instead of the full toolset when an agent needs one capability.
• Constrain generic tools: load_document that validates URLs instead of open-ended fetch_url.
• tool_choice: auto (default), any, tool + name, none; force the first step when order matters.
• Current flagship models reject any and tool with a 400; use auto with strict: true or structured outputs.
• disable_parallel_tool_use: true, set inside tool_choice, limits Claude to one tool call per turn.

Worked example: An insurance claims pipeline

Intake, assessment and payout agents all shared one 16-tool list, and the assessment agent sometimes triggered payments. Giving each agent only its role's tools (intake: read_claim, request_documents; assessment: read_claim, policy_lookup; payout: issue_payment behind an approval check) removed the misuse and made tool choice easier.

Exam traps

• Giving every agent every tool 'for flexibility'.
• Fixing a latency problem by over-provisioning the synthesis agent with all search tools.
• Assuming tool_choice any or tool works on every model; current flagships return 400.
• Using an open-ended fetch tool where a validated, purpose-specific one would do.

Official docs

• Define tools (tool_choice)
• Writing effective tools for agents

Practise CCAR-F questions · Next lesson: MCP servers in Claude Code and agents