Every assignment is a real system, not a textbook exercise. You push to GitHub, you demo
it live, you own it.
Capstone Project
Production Multi-Agent System
Your capstone is a complete multi-agent system built with the Google Agent Development
Kit (ADK). You design the agent topology — a coordinator delegating to specialised sub-agents — give each
agent its own tools, MCP connections, and Agent Skills, wire in short-term session state and long-term
memory, add guardrails and evaluation, then deploy the whole system to the Google ecosystem with a live,
shareable endpoint. You choose the domain; the architecture is what you're graded on.
Google ADK
Multi-Agent Orchestration
Agent Tools
MCP Servers
Agent Skills
Session State & Memory
Vertex AI Agent Engine
Cloud Run
Gemini
LangSmith
LangFuse
Guardrails
Week 1–2 Lab
ReAct Agent from scratch
Build the Thought → Action → Observation loop manually in Python — no framework.
Understand exactly how the stop token prevents hallucinated observations and why this loop is the foundation
of all agent frameworks.
OpenAI API
Python
JSON Schema
Week 3 Lab
RAG Research Agent
A LangChain agent with three tools: RAG over a custom document set, a Python REPL for
live calculation, and Wikipedia search. Produces structured answers with source citations — with every run
traced end-to-end in LangSmith and LangFuse so you can see exactly which tool call cost what.
LangChain LCEL
Google ADK
LangSmith
LangFuse
Week 4–5 Lab
Multi-Agent Content Pipeline
A LangGraph pipeline with Researcher, Writer, and Editor agents supervised by an
orchestrator. Includes a human-in-the-loop approval gate that pauses execution, waits for review, and
resumes — fully checkpointed.
LangGraph StateGraph
MemorySaver
HITL
MCP Integration
Custom MCP Server
Build and deploy your own Model Context Protocol server in Python — then connect it to
a LangGraph agent as a tool node and to a Google ADK agent as an MCP toolset. Write the integration once;
use it across every framework.
MCP Protocol
stdio transport
HTTP+SSE
Google ADK
Memory Lab
Agent with persistent memory
Give an agent a memory that survives the session. Implement short-term session state,
long-term vector memory with semantic recall, and summarisation to keep context windows under control — then
watch the same agent answer a follow-up question three days later.
Session State
Vector Memory
Embeddings
Context Compaction
Skills Lab
Reusable Agent Skills library
Package agent capabilities as self-contained Skills — instructions, tools, and
resources bundled together — so a capability you write once can be loaded by any agent in your multi-agent
system instead of copy-pasted into every prompt.
Agent Skills
Tool Definitions
Google ADK
