Session 3: Agents - from pipeline to system that decides
14 lessons · Back to full syllabus
What you keep
What an agent really is, how to build the loop by hand, and how to ship one on Google ADK (with LangGraph as an accepted alternative for explicit graph control).
You ship
Your assistant turned into an agent that plans and uses tools, plus a Streamlit UI that demos the loop.
Recommended reading (async)
Multi-agent foundations - watch & read (async)
Before or after the live Session 3 deck, these videos and docs give you production multi-agent patterns and the Claude Agent SDK as a programmable harness - complementary to ADK / LangGraph hands-on work in the lessons below.
Videos
The Multi-Agent Architecture That Actually Ships
Luke Alvoeiro (Factory) at AI Engineer - orchestrators, workers, validators, and validation contracts that hold up in production.
Multi Agent Systems Explained: How AI Agents & LLMs Work Together
IBM Technology - plain-English intro to multi-agent cooperation, structures, and how LLMs + tools fit together.
Live session resources
Week 3 session: Agents with ADK
Agent loop, MCP, multi-agent, and A2A. Live session deck for Engineering Week 3.
Open resourceWeek 3 session: Agents with ADK — extended cut
Same live session deck, plus a bonus "New Kids on the Block" section on graph, loop, harness, and context engineering.
Open resourceWeek 3 assignment build guide
Step-by-step Path A / Path B guide - Google ADK preferred, LangGraph accepted - with Streamlit UI required and copy-paste prompts.
Open resourceClaude Agent SDK overview
Build production agents with Claude Code as a library - same tools, agent loop, and context management in Python and TypeScript.
Open resourceADK sample code (preferred)
Routing, MCP, and full-system demos with Google ADK. Clone, run Demo 1, then adapt to your capstone.
Open resourceLangGraph sample code (accepted alternative)
Same three demos as the ADK repo, re-implemented in LangGraph. Use if ADK is blocked or you prefer graphs.
Open resourceGoogle ADK docs
ADK agents, tools, MCP, multi-agent, and A2A (default assignment stack).
Open resourceLangGraph docs + Academy
Accepted alternative stack: overview, quickstart, persistence, and free Intro to LangGraph course.
Open resourceTool call orchestration
Parallel vs sequential tool calls, defensive result parsing, and turning tool failures into self-healing retries. The part frameworks hide from you.
Open resourcePrompt chain vs pipeline vs graph vs agent
The cost-benefit of an agent loop in tokens and latency - and when a single prompt, a fixed chain, or a deterministic graph beats building an agent at all. Real examples of over-engineered agents.
Open resourceLong context engineering
Why agents rot: context rot across 18 frontier LLMs, the n² attention budget, the U-shaped recall curve - and the four pillars (Write, Select, Compress, Isolate) that keep decision-relevant context in the window.
Open resourceLessons
The agent loop, by hand with the raw SDK
Think, pick a tool, call it, observe, decide again - no framework mystification.
Orchestration with LangGraph
Nodes, edges, shared state - durable execution and human-in-the-loop checkpoints.
Tools and MCP
MCP is how modern agents get tools - connect to servers instead of hand-wiring every integration.
Multi-agent and A2A
Several specialists when roles are genuinely separable - most problems do not need multiple agents.
Prompt injection, where guardrails get dangerous
Malicious instructions in retrieved content can hijack tool-using agents.
Long context engineering: why agents rot, and the four pillars that keep them sharp
Context rot, the U-shaped recall curve, and the Write · Select · Compress · Isolate moves that keep an agent sharp from turn 10 to turn 500.
Build your own MCP server
Tools, resources, and prompts, then a real Python FastMCP server you can wire into Claude Code today.
Build your own A2A integration
Agent Card, skills, and tasks - a minimal A2A server and the client call that reaches it.
LangGraph in depth
Reducers, Command routing, subgraphs, checkpointers, time travel - what the live session does not have time for.
Agent security mitigations
Least privilege, human-in-the-loop, structural separation of instructions from content, sandboxing, and an audit trail.
Tool call orchestration: parallel calls, result parsing, and error recovery
Parallel vs sequential tool calls, defensive result parsing, and teaching the model to recover from failures instead of crashing.
Prompt chain, pipeline, graph, or agent? Knowing when a loop is worth it
The cost-benefit of an agent loop in tokens and latency - and when a single prompt, a chain, or a deterministic graph beats building an agent at all.
Session 3 assignment support: Turn your capstone into an agent
Step-by-step: pick the job, build a Google ADK agent with one real tool and a Streamlit UI, prove Think → Act → Observe, then optional stretch (LangGraph accepted as alt).
Turn your capstone into an agent
Google ADK agent with at least one real tool and a Streamlit UI - trace each step of the decision loop. LangGraph accepted as alternative. Stretch: HITL, MCP, A2A.
Lessons in this module