CCA-F: Claude Certified Architect — Foundations

Question

You are building a customer support resolution agent using the Claude Agent SDK. The agent handles high-ambiguity requests such as returns, billing disputes, and account issues, and has access to backend systems through custom Model Context Protocol (MCP) tools: get_customer, lookup_order, process_refund, and escalate_to_human. Your target is 80%+ first-contact resolution while still knowing when to escalate.

After expanding the agent's MCP tools with delivery-specific capabilities — check_delivery_status, contact_driver, issue_credit, apply_promo_code, update_delivery_address, and reschedule_delivery — the total tool count grows from 4 to 10. Your evaluation suite shows tool selection accuracy dropping from 88% to 71%. Log analysis reveals that most errors involve the agent choosing between semantically overlapping tools: calling issue_credit when process_refund was the correct choice, and calling check_delivery_status when lookup_order already returns the needed data.

Which approach structurally eliminates the semantic overlap identified in the logs as the source of these errors?

Explanation

The errors stem from having multiple distinct tools that can plausibly handle the same underlying action, forcing the agent to guess between them. Enabling deferred tool loading, splitting tools across sub-agents, or adding few-shot examples all work around this ambiguity without removing it — the overlapping tools still exist as separate, competing options whenever they are visible to the model. Merging issue_credit and process_refund into a single resolve_compensation tool (with an action parameter) and folding check_delivery_status into lookup_order (via an include_tracking flag) removes the redundant tool definitions entirely, so there is no longer a choice to make between two similar tools — the agent selects one tool and expresses intent through parameters. This directly addresses the documented best practice for agentic tool design of consolidating overlapping functionality to reduce tool-selection ambiguity, rather than compensating for it through routing, visibility control, or prompting.

CCA-F: Claude Certified Architect — Foundations

AstroTutor

Notes

Readiness

QuestionQ55

Community Discussion