Agentic AI Architecture

This instructor-led course teaches experienced developers how to design, build, and deploy production-ready multi-agent AI systems. The program combines hands-on implementation with architectural best practices, focusing on tool integration, retrieval-augmented generation, orchestration patterns, and enterprise-grade concerns like observability, evaluation, and security.

Participants build a complete Event Operations Planner — a multi-agent system that coordinates specialized AI agents to plan safe, compliant events with human-in-the-loop approvals.

Learning Objectives

By completing this course, participants will be able to:

• Architect multi-agent systems using the supervisor orchestration pattern
• Engineer tool-calling agents using the Model Context Protocol (MCP)
• Build RAG systems with semantic search for policy grounding
• Orchestrate agent workflows with LangGraph state machines
• Add structured observability and distributed tracing to agentic systems
• Create human-in-the-loop approval workflows with revision loops
• Evaluate agent behavior using functional, RAG, and LLM-as-judge evaluators
• Secure agents against prompt injection and implement defense-in-depth

Course Modules

Day 1: Foundations

Lab 1: Tools & MCP — Building the foundation for agent-world interaction through FastAPI endpoints and the Model Context Protocol. Students implement venue and weather APIs, then wrap them as MCP tools that agents can discover and invoke.

Lab 2: Retrieval-Augmented Generation — Grounding agent decisions in enterprise knowledge. Students build a complete RAG pipeline: document chunking, FAISS vector indexing, semantic search, and retrieval formatting for LLM consumption.

Lab 3: Single Agent Anatomy — Understanding the core agent loop pattern. Students implement a Policy/RAG Agent that searches the knowledge base, extracts constraints with citations, and demonstrates the "retrieve or refuse" pattern for grounded behavior.

Day 2: Scale

Lab 4: Multi-Agent Orchestration — Coordinating multiple specialist agents using LangGraph. Students build a supervisor that routes between Intake, Policy, Planner, and Approvals agents based on a status-driven state machine.

Lab 5: Observability — Understanding agent behavior in production. Students add structured logging with correlation IDs, implement audit trails, and integrate distributed tracing for end-to-end visibility.

Lab 6: Human-in-the-Loop — Knowing when and how to involve humans. Students implement dual-approval workflows requiring both Security Lead and Operations Lead sign-off, with revision loops that route rejected plans back for improvement.

Day 3: Production

Lab 7: Evaluation — Testing agentic systems systematically. Students build an evaluation framework with functional evaluators (status checks, constraint validation), RAG evaluators (faithfulness, context precision), and LLM-as-judge evaluators for subjective quality assessment.

Lab 8: Security & Governance — Protecting agentic systems from adversarial inputs. Students implement layered prompt injection detection using pattern matching and risk scoring, configure block vs. log-only modes, and apply defense-in-depth principles.

Technology Stack

Capstone Project

The Event Operations Planner capstone brings all course concepts together in a realistic, production-grade agentic system. Participants build a multi-agent application that coordinates intake, policy discovery, risk assessment, planning, and approvals to produce safe, compliant event plans. The system integrates real tools via MCP, grounds decisions using retrieval-augmented generation, routes work through a supervisor-driven workflow, and enforces human-in-the-loop approvals from security and operations leaders. By the end of the course, learners have implemented a fully observable, evaluated, and secured agentic system that mirrors the challenges of deploying AI in real enterprise environments.

Audience

Python developers and ML engineers interested in building production-grade multi-agent AI systems. Ideal for teams planning to deploy agentic applications in enterprise environments where reliability, observability, and security are critical.

Materials Provided

• Complete lab exercises with scaffolded code
• Reference implementations (checkpoints) for each lab
• Knowledge base documents (policies, templates, postmortems)
• Evaluation scenarios and test suites
• Full capstone solution with Streamlit UIs

• Developers comfortable scripting and building automation in Python (or similar languages)
• Intermediate knowledge of concpets such as async/await, classes, type hints
• Basic understanding of REST APIs
• Familiarity with LLMs and prompt engineering concepts
• Development environment with Python 3.11+, Poetry, and an LLM API key (OpenAI or Anthropic)