Customer Support AI Agent — Amazon Bedrock AgentCore Project

Udacity AWS AI Engineering Nanodegree, Course 2.

This is my submission for the customer support agent project. I built and deployed the agent described in the project instructions using Amazon Bedrock AgentCore, the Strands SDK, and a handful of AWS services (Lambda, API Gateway, Bedrock Knowledge Bases, AgentCore Memory, Code Interpreter, and the AgentCore Browser tool).

Agent ARN: arn:aws:bedrock-agentcore:us-east-1:092134045103:runtime/customer_support_agent-BVDt7HAJnn

What it does

• Tracks orders and processes refunds through Lambda functions wired up as Gateway tools (MCP)
• Answers product/policy questions using a Bedrock Knowledge Base (RAG)
• Remembers customer name and preferences across separate sessions
• Calculates loyalty discounts with exact math using the Code Interpreter
• Can browse a live web page and read back what it finds

Task status

Task	Done?
1. App init + clients (TODO 1-3)	Yes
2. Namespace helper (TODO 4)	Yes
3. MemoryHook class (TODO 5)	Yes
4. Knowledge Base search tool (TODO 6)	Yes
5. Loyalty discount calculator (TODO 7)	Yes
6. Agent entrypoint (TODO 8)	Yes
7. Deploy + run all 6 test scenarios	Yes
8. Written reflection	Yes

Test results (real command + real output for all 6 scenarios): test_logs/test_results.md

Reflection: REFLECTION.md

Extra stuff beyond the base requirements

The project suggests a few optional additions to go further. I did two of them:

• Structured output validation. calculate_loyalty_discount now returns results validated against a Pydantic model (DiscountResult) with all the required fields (points_redeemed, tier_discount_pct, final_total, remaining_points, plus total_savings/points_earned/note). This also fixed a bug I found while doing it — the tool used to return the Code Interpreter's raw response wrapper instead of the actual numbers.
• Conversation summarization. The agent now uses Strands' SummarizingConversationManager so that if a single request chains a lot of tool calls (the browser tool especially can take several steps), older messages get summarized instead of just dropped or blindly kept, which keeps token usage under control.

I did the third one too (domain personalization), but kept it as a totally separate deployment so it doesn't touch this graded submission at all: hotel_variant/ is the same agent re-themed for a hotel — reservations instead of orders, cancellations instead of refunds, guest rewards instead of loyalty discount. Same architecture, own Lambdas/Gateway/Knowledge Base/Memory/Runtime, own test results in hotel_variant/TEST_RESULTS.md. While testing that one I actually caught something interesting — see the note at the bottom of this README.

Project structure

project/
├── README.md
├── REFLECTION.md
├── test_logs/
│   ├── test_results.md
│   └── test*.png            (screenshots for each test)
├── starter/                  (the graded submission)
│   ├── main.py                (the completed agent)
│   ├── pyproject.toml
│   ├── product_catalog.txt    (uploaded to S3, synced into the Knowledge Base)
│   └── lambda/
│       ├── order_tracker.py
│       ├── refund_processor.py
│       └── lambda_schema
└── hotel_variant/             (bonus: same agent, hotel domain, separate deployment)
    ├── README.md
    ├── TEST_RESULTS.md
    ├── main.py
    ├── pyproject.toml
    ├── hotel_policies.txt
    └── lambda/
        ├── reservation_tracker.py
        ├── cancellation_processor.py
        └── lambda_schema

AWS resources I actually deployed

Resource	Value
Region	us-east-1
Lambda (orders)	order-tracker
Lambda (refunds)	refund-processor
API Gateway REST API	chcb0zkb9b, stage prod
AgentCore Gateway	customersupportgateway-zh3m74vmjj, NONE authorizer, 2 targets (order-tracker API Gateway target + refund-processor Lambda target), 6 MCP tools total
Knowledge Base	CustomerSupportKB, ID USCGD9ZEJ1
AgentCore Memory	CustomerSupportMemory-L1eStICBN4, strategies customer_facts (semantic) and customer_preferences (user preference)
AgentCore Runtime	customer_support_agent-BVDt7HAJnn, Direct Code Deploy, Python 3.11

Model used: Amazon Nova Lite (global.amazon.nova-2-lite-v1:0).

Checking Gateway tools:

npx @modelcontextprotocol/inspector
# connect to the Gateway URL, should list 6 tools

Checking the Knowledge Base directly:

aws bedrock-agent-runtime retrieve \
  --knowledge-base-id USCGD9ZEJ1 \
  --retrieval-query '{"text": "What is the return policy for electronics?"}' \
  --region us-east-1
# should mention the 15-day electronics return window

Local setup / deploy

cd starter
uv sync
AGENTCORE_SUPPRESS_RECOMMENDATION=1 agentcore configure --entrypoint main.py --name customer_support_agent
AGENTCORE_SUPPRESS_RECOMMENDATION=1 agentcore deploy

Running the 6 test scenarios yourself

These commands are copy-pasteable. Run them one at a time from the starter/ folder. $(uuidgen) just generates a fresh session id each time — AgentCore needs the --session-id flag to be at least 33 characters.

Test 1 — Order Tracking

AGENTCORE_SUPPRESS_RECOMMENDATION=1 agentcore invoke --session-id "$(uuidgen)" \
  '{"prompt": "Can you track order ORD-001?", "customer_id": "CUST-123", "session_id": "t1"}'

Test 2 — Refund Processing

AGENTCORE_SUPPRESS_RECOMMENDATION=1 agentcore invoke --session-id "$(uuidgen)" \
  '{"prompt": "I want to return my Kindle Paperwhite (ORD-002). Please initiate a refund.", "customer_id": "CUST-123", "session_id": "t2"}'

Test 3 — Knowledge Base (RAG)

AGENTCORE_SUPPRESS_RECOMMENDATION=1 agentcore invoke --session-id "$(uuidgen)" \
  '{"prompt": "What are the benefits of the Platinum loyalty tier?", "customer_id": "CUST-123", "session_id": "t3"}'

Test 4 — Long-Term Memory (needs two calls, two screenshots)

Session A, introduce yourself:

AGENTCORE_SUPPRESS_RECOMMENDATION=1 agentcore invoke --session-id "$(uuidgen)" \
  '{"prompt": "Hi, I am Nada Feteiha. I prefer concise responses.", "customer_id": "CUST-DEMO2", "session_id": "s-A"}'

Wait about a minute for memory extraction to run, then Session B (brand new session, same customer_id), to check recall:

AGENTCORE_SUPPRESS_RECOMMENDATION=1 agentcore invoke --session-id "$(uuidgen)" \
  '{"prompt": "Do you remember my name and communication preference?", "customer_id": "CUST-DEMO2", "session_id": "s-B"}'

Test 5 — Loyalty Discount Calculation

AGENTCORE_SUPPRESS_RECOMMENDATION=1 agentcore invoke --session-id "$(uuidgen)" \
  '{"prompt": "I am a Gold member with 4250 points. Calculate my discount on a $150 standard order.", "customer_id": "CUST-123", "session_id": "t5"}'

Test 6 — Browser Tool

AGENTCORE_SUPPRESS_RECOMMENDATION=1 agentcore invoke --session-id "$(uuidgen)" \
  '{"prompt": "Go to https://www.udacity.com and tell me the page title.", "customer_id": "CUST-123", "session_id": "t6"}'

(This one takes a bit longer, the browser session has to spin up first.)

Notes / gotchas I ran into

• The AgentCore CLI's own --session-id needs to be 33+ characters, otherwise it errors out. I use uuidgen for that.
• If you reuse the same customer_id for a lot of testing, the memory system starts injecting older, unrelated facts into new conversations because retrieve_customer_context has no relevance threshold. Use a fresh customer_id when you want a clean memory demo.
• The browser tool needs a session_name matching ^[a-z0-9-]+$, 10-36 characters. I added a line in the system prompt telling the model this explicitly, otherwise it sometimes picked an invalid name and just gave up after the first error instead of retrying.
• More detail on both of these is in REFLECTION.md.
• While testing the discount calculation on the hotel variant, one run came back with the wrong final numbers even though I'd already verified the tool itself always computes correctly. A retry on the exact same prompt gave the right answer. Looks like Nova Lite occasionally messes up when turning a correct tool result into a sentence, rather than an actual bug — see hotel_variant/TEST_RESULTS.md for the details.

References

• Amazon Bedrock AgentCore Documentation
• Strands Agents Documentation
• MCP Inspector
• uv Package Manager