Skip to main content

​
Project Goal

Build a customer support agent that demonstrates the three core concepts from this module:
  • ✅ Agent basics - Tool use with proper agent loop
  • ✅ Well-designed tools - Following MCP best practices
  • ✅ Memory management - Working memory for conversation + long-term for preferences

​
Architecture

User Query
    |
    V
+-------------------------------------+
│ Customer Support Agent              │
│ - Maintains conversation context    │
│ - Remembers user preferences        │
│ - Uses tools intelligently          │
+-------------------------------------+
          |
          |--> [search_knowledge_base]
          |--> [get_customer_info]
          |--> [check_order_status]
          |--> [create_support_ticket]
          |
          V
    Working Memory (session)
          +
  Long-Term Memory (preferences)

​
Project Requirements

1. Agent Foundation Must implement:
  • Basic agent loop with tool use
  • At least 3 tools from customer support domain
  • Proper error handling for tool failures
  • Graceful responses when no tool is needed
2. Tool Design (MCP Best Practices) Each tool must have:
  • Clear, descriptive name (verb_noun_context format)
  • Comprehensive description (what, when to use, when NOT to use)
  • Simple parameter schema (3 or fewer parameters preferred)
  • Consistent response format (success/error envelope)
  • Graceful error handling with actionable messages
Recommended tools:
  • search_knowledge_base(query, category) - Find help articles
  • get_customer_info(email) - Look up customer account
  • check_order_status(order_id) - Track order/delivery
  • create_support_ticket(email, subject, description) - Escalate to human
3. Memory Implementation Must implement:
  • Working memory for conversation continuity
  • Tool result caching (avoid redundant calls within session)
  • Long-term memory for user preferences
  • Choose one integration pattern (code-driven or background extraction)
Example preferences to track:
  • Communication style (formal/casual)
  • Preferred contact method (email/phone)
  • Product interests
  • Past issues and resolutions
4. Testing Requirements Must demonstrate:
  • Multi-turn conversation with context retention
  • Tool selection accuracy (right tool for each query)
  • Memory retrieval (reference previous conversation)
  • Error handling (tool failure gracefully handled)
  • Preference learning and application
Test scenarios:
  1. Order tracking: “I ordered a laptop last week” → “When will it arrive?” (should remember order)
  2. Knowledge base: “How do I reset my password?” (search articles)
  3. Escalation: “This doesn’t work, I need help” (create ticket)
  4. Preference: “I prefer email communication” → later: “Contact me about this” (should use email)

​
Bonus Challenges

Choose one or more:
  • Tool consolidation: Combine related data fetches into single efficient tool
  • Semantic enrichment: Add contextual insights to tool responses
  • Custom memory strategy: Implement domain-specific extraction
  • Multiple patterns: Use both code-driven and LLM-driven memory
  • Advanced caching: Implement intelligent cache invalidation

​
Resources