🇫🇷 Version française

[!IMPORTANT] Every fixture, rulebook, and calculator output in this lab is synthetic and non-binding. Nothing here represents an actual Desjardins product, rate, or policy, and no regulator or insurer has reviewed or endorsed this material.

Overview

Item	Value
Duration	35 minutes
Level	Advanced
Prerequisites	Lab 02, Lab 03, Lab 04, Lab 05

Learning Objectives

By the end of this lab, you will be able to:

• Name the agent’s three sequential stages: intake, reference lookup, composition
• Explain the supervisor’s routing rule in decide_next_step, including the short-circuit for an invalid case reference
• Explain why toolbox.py calls the Lab 04/05 MCP tool functions in-process instead of opening a network client session
• Confirm the agent’s own code never calls approve, reject, or revise on the approval repository

Exercises

Exercise 6.1: Read the Supervisor’s Routing Logic

Open src/quote-preparation-agent/graph.py and read decide_next_step. The supervisor always visits intake first. If intake found the case reference valid, it routes to reference_lookup next; if intake found it invalid, it skips straight to composition so composition can still produce a bounded rejection message. Once composition completes, the graph ends.

Exercise 6.2: Read Why Toolbox Calls Are In-Process

Open src/quote-preparation-agent/toolbox.py and read its module docstring. It explains that this phase calls get_application and get_rulebook by loading each MCP server’s main.py directly, rather than opening an mcp Python SDK ClientSession against a running server. This keeps tests fast and free of any open port, while still exercising the exact functions each server exposes over MCP. Note also that approve, reject, revise, and open_training_preview are intentionally never imported here: only a human reviewer may call those.

Exercise 6.3: Run the Existing Tests

python -m pytest src/quote-preparation-agent/tests/test_toolbox.py -v

Expected result: every test passes, confirming the toolbox wrappers call through to the calculator, the approval repository, and both MCP servers correctly.

Exercise 6.4 (Hands-on): Inject a Stub Model

build_graph in graph.py accepts an optional model callable, defaulting to default_model, which makes no network call. Build the graph with your own stub and confirm the intake node’s note reflects it.

python -c "
import sys
sys.path.insert(0, 'src/quote-preparation-agent')
from graph import build_graph

def my_stub_model(prompt: str) -> str:
    return f'[lab-06-stub] {prompt}'

graph = build_graph(model=my_stub_model)
result = graph.invoke({
    'case_id': 'CASE-SYN-001',
    'preparer_id': 'AGENT-INTAKE',
    'rulebook_id': 'RULEBOOK-SYN-ON',
    'intake_complete': False,
    'lookup_complete': False,
    'composition_complete': False,
})
print(result['intake_note'])
"

Expected result: the printed note starts with [lab-06-stub], confirming the model callable is a genuine seam and not hardcoded.

Validation Checklist

• pytest src/quote-preparation-agent/tests/test_toolbox.py -v passes
• You can name the three sequential stages and explain the invalid-case-reference short-circuit
• You located the toolbox docstring explaining the in-process MCP call design choice
• Your Exercise 6.4 stub model’s text appeared in the graph’s output

Knowledge Check

• Why does toolbox.py never import approve, reject, revise, or open_training_preview from the approval repository?
• What would have to change in toolbox.py for this agent to call a deployed MCP server instead of an in-process function?

Next Steps

Continue to Lab 07: Run the Agent End to End.