Skip to content

COURSE-PLAN.md

Latest commit

 

History

History
77 lines (60 loc) · 3.49 KB

COURSE-PLAN.md

File metadata and controls

77 lines (60 loc) · 3.49 KB

Build Production-Ready AI Agents — Course Plan

O'Reilly Live Learning | 4 × 50-minute segments + 10-minute breaks

Segment 1 — Understanding Agents at Depth (50 min)

Theme: What makes an agent an agent, and how do the pieces fit together?

  • Agent anatomy: perception → reasoning → memory → action loop
  • Token limits, context windows, and why they constrain agent design
  • Memory types: in-context, external (vector/SQL), episodic, semantic
  • Tool use and function calling — how agents reach outside the LLM
  • MCP (Model Context Protocol) — what it is and why it matters
  • Agent patterns: ReAct, plan-and-execute, multi-agent orchestration
  • Demo: Claude Code custom agents — show a subagent being dispatched
  • Demo: Claude Code skills — invoke a skill, show the prompt expansion
  • Demo: Claude Code CLAUDE.md / agent context files
  • Q&A / break

Segment 2 — Low-Code Agent Deep-Dive (50 min)

Theme: Copilot Studio as an opinionated, visual agent runtime

  • Copilot Studio orientation: topics, actions, knowledge, variables
  • Build a Contoso HR Agent in Copilot Studio (no-code version)
    • Add HR policy knowledge source (SharePoint / URL)
    • Create a resume intake topic with triggers and entities
    • Wire a connector action (e.g. send email on decision)
    • Add generative answers node grounded in policy docs
  • Global variables and conversation state
  • Adaptive Cards for structured output
  • Publish to Teams channel — end-to-end demo
  • Compare low-code vs code-first: when to use each
  • Q&A / break

Segment 3 — Code-First Agentic AI (50 min)

Theme: The Contoso HR Agent — LangGraph + CrewAI in production patterns

  • Repo tour: contoso-hr-agent/ structure, uv workflow, .env setup
  • LangGraph fundamentals: StateGraph, nodes, edges, checkpointing
  • Key demo: Parallel fan-out — policy_expert ‖ resume_analyst run concurrently
    • Show the graph wiring (add_edge fan-out / fan-in)
    • Show partial state returns — why parallel nodes can't do {**state, ...}
  • CrewAI agents: persona, tools, Crew.kickoff() per node
  • ChromaDB RAG: vectorizer → retriever → query_hr_policy tool
  • Chat memory: localStorage + server JSON + cross-session context injection
  • Live run: Drop a resume → watch all five pipeline nodes fire in terminal
  • Pipeline Runs page (runs.html): walk through the trace together
  • MCP server preview: uv run hr-mcp, hit tools in MCP Inspector
  • Q&A / break

Segment 4 — Deployment, Monitoring, Optimization (50 min)

Theme: Getting agents to production and keeping them there

  • Azure deployment patterns for agents (ACA, App Service, AKS)
  • Azure AI Foundry: model deployment, endpoint management, API versioning
  • Observability: LangSmith tracing, Application Insights, structured logging
  • MCP in the real world
    • Connect MCP server to Claude Code (claude_desktop_config.json)
    • Connect MCP server to GitHub Copilot in VS Code
    • MCP Inspector for live tool testing and debugging
  • Cost and token optimization: model routing, caching, prompt compression
  • Security: secret management (Key Vault), input validation, PII handling
  • Evaluation and regression testing for agent outputs
  • Shavings sweep: common failure modes, retry patterns, graceful degradation
  • Resources, next steps, community links
  • Final Q&A