Running the capstone
Let's run the full Office Manager demo and watch all the concepts work together.
# Create and run the Office Manager
manager = OfficeManager(DEFAULT_MODEL)
manager.run(
"Reply to my Boss saying 'Yes, I'll be there' "
"and schedule a meeting for Friday 2pm."
)
# Expected flow:
# 1. LLM calls send_email(to="Boss", body="Yes...")
# 2. HITL: "Approve send_email to Boss?" -> yes
# 3. Email sent!
# 4. LLM calls schedule_meeting(time="Friday 2pm")
# 5. Meeting scheduled (auto-execute, safe tool)
# 6. LLM: "Done! Email sent and meeting scheduled."The full scenario: LLM decides on tools, HITL gate for email, auto-execute for calendar, final response.
Validation checklist: Capstone verification checklist
Loading practice…
Look at what you have built across this course: 1. Stateful agents with conversation memory 2. ReAct reasoning with thought-action-observation 3. Native function calling with JSON schemas 4. Autonomous multi-step loops with safety limits 5. Sliding window and summarization memory 6. Router and chaining workflow patterns 7. Validation and self-correction with reflection 8. Human-in-the-loop approval gates 9. A complete Office Manager integrating everything All from scratch, no frameworks, pure Python + LiteLLM.
Where to go from here: you now understand the foundations that every agent framework builds on. You can confidently use LangChain, CrewAI, or any other framework because you know what happens under the hood. Or build your own custom agents using these exact patterns.