Assignment

Turn your capstone into an agent

Tip

Need step-by-step support?

Use the full build guide here: Turn your capstone into an agent.

Ready when

Default stack: Google ADK

Preferred: Google ADK + at least one real tool (matches the Week 3 live session). Accepted alternative: LangGraph or a clearly traced raw-SDK loop - same proof bar.

Tip

Sample code (start here)

ADK: adk-multi-agent-systems. LangGraph twin: langgraph-multi-agent-systems. Clone, run Demo 1, adapt to your capstone - do not submit the stock demo.

Tip

Session deck + docs

Week 3 session: /courses/ai-engineering-bootcamp/week-3-agents. ADK: docs · agents · MCP tools. Alt: LangGraph.

Rebuild your capstone assistant as an agent that plans and acts, not only answers in one shot. Default stack for the assignment: Google ADK + at least one real tool + a Streamlit UI that demos the loop. Start from the ADK sample repo, then adapt. Optional stretch: multi-agent / A2A, MCP, HITL, or a LangGraph mirror of the same job.

Part 1 - Required

  • A Google ADK agent (preferred) - or LangGraph / clearly traced raw-SDK - that completes a multi-step task using a tool.
  • You can show each step of the decision loop (Think / Act / Observe) from logs or a short Loom.
  • A Streamlit UI that runs or demos the agent (screenshot or short Loom).
  • A one-liner: why this is an agent (or workflow) and why that choice is right.

Part 2 - Stretch

  • Human-in-the-loop before one consequential action.
  • ADK multi-agent / router / A2A (only if roles are truly separable).
  • One MCP-backed tool (natural fit with ADK after the live session).
  • Mirror the same job in LangGraph (if you shipped ADK) - or ADK if you shipped LangGraph.
  • A deliberate prompt-injection test case and how you defended against it.

Submit

Post in Maven / WhatsApp: short clip or screenshots of the loop (including Streamlit UI), stack used (ADK / LangGraph / raw), link or export of the agent/code, and which stretch items you hit. Do not paste secrets.

Done when

  • It completes a multi-step task using a tool
  • You can trace each step of its decision loop
  • Streamlit UI demos the agent (screenshot or short Loom)
  • You can say in one line whether it is a workflow or an agent and why

Submit in Maven: Clip of the loop running (include Streamlit UI), how you built it (ADK preferred / LangGraph / raw), and any stretch goals.