Advanced AI Agents Masterclass: Engineering, Security, and Governance for Production Systems Training Course

Day 1

Anatomy of a Modern AI Agent

Agents as autonomous reasoning and action systems, extending beyond basic chatbots

Paradigms of reactive, proactive, hybrid, and goal-directed agents

Core components: perception, planning, memory, tool utilization, and action

Trade-offs between single-agent and multi-agent design strategies

Agent Frameworks and the Modern Stack

Analysis of LangChain, LlamaIndex, AutoGen, and CrewAI, highlighting their respective trade-offs

Comparison with classical frameworks such as JADE and SPADE

Selecting frameworks based on specific production requirements

Implementing tool calling, function calling, and structured outputs

Hands-on: Scaffolding a single Python agent with tool calls

Multi-Agent System Architectures

Designs for centralized, decentralized, hybrid, and layered Multi-Agent Systems (MAS)

FIPA ACL, message-passing, and their modern equivalents

Coordination patterns including planning, negotiation, and synchronization

Emergent behaviors and self-organization within agent populations

Decision-Making and Learning in Agents

Applying game theory to cooperative and competitive agent interactions

Reinforcement learning in multi-agent environments

Transfer learning and knowledge sharing across agents

Conflict resolution and establishing trust among coordinating agents

Day 2

Multi-Modal Foundations for Agents

Multi-modal AI as a unified workflow integrating text, image, speech, and video

Leading multi-modal models: GPT-4 Vision, Gemini, Claude, and Whisper

Fusion techniques for combining modalities within an agent's reasoning loop

Trade-offs regarding latency, cost, and accuracy in multi-modal pipelines

Building the Perception Layer

Image processing for agents: classification, captioning, and object detection

Speech recognition using Whisper ASR and streaming transcription

Text-to-speech synthesis and natural voice interaction

Linking perception outputs to LLM-driven reasoning and tool selection

Hands-On - Building a Multi-Modal Agent in Python

Defining the agent's task, context window, and tool inventory

Integrating GPT-4 Vision and Whisper APIs end-to-end

Implementing memory, state, and conversation management

Adding tool calls that safely produce real-world side effects

Hands-On - Orchestrating a Multi-Agent System

Composing specialized agents using AutoGen or CrewAI

Defining roles, responsibilities, and inter-agent communication protocols

Resource allocation and coordination in a simulated environment

Logging agent reasoning, tool calls, and decisions for inspection and audit

Day 3

Threat Surface of Production AI Agents

Understanding what makes agentic AI uniquely vulnerable compared to traditional software

Attack surfaces: data, model, prompt, tool, output, and interface layers

Threat modeling for agent-based systems with autonomous tool usage

Comparing AI cybersecurity practices with traditional cybersecurity

Adversarial Attacks Hands-On

Adversarial examples and perturbation methods: FGSM, PGD, and DeepFool

White-box versus black-box attack scenarios

Model inversion and membership inference attacks

Data poisoning and backdoor injection during training

Prompt injection, jailbreaking, and tool misuse in LLM-based agents

Defensive Techniques and Model Hardening

Adversarial training and data augmentation strategies

Defensive distillation and other robustness techniques

Input preprocessing, gradient masking, and regularization

Differential privacy, noise injection, and privacy budgets

Federated learning and secure aggregation for distributed training

Hands-On with the Adversarial Robustness Toolbox

Simulating attacks against the multi-modal agent built on Day 2

Measuring robustness under perturbation and quantifying degradation

Iteratively applying defenses and re-evaluating attack success rates

Stress-testing tool-call pathways and prompt injection vectors

Day 4

Risk Management Frameworks for AI

NIST AI Risk Management Framework: govern, map, measure, manage

ISO/IEC 42001 and emerging AI-specific standards

Mapping AI risk to existing enterprise GRC frameworks

Requirements for AI accountability, auditability, and documentation

Regulatory Compliance for Agentic Systems

EU AI Act: risk tiers, prohibited uses, and obligations for high-risk systems

GDPR and CCPA implications for agent data pipelines

U.S. Executive Order on Safe, Secure, and Trustworthy AI

Sector-specific guidance for finance, healthcare, and public services

Third-party risk and supplier AI tool usage

Ethics, Bias, and Explainability

Bias detection and mitigation across agent perception and reasoning

Explainability and transparency as security-relevant properties

Fairness, downstream harm, and responsible deployment

Designing inclusive and auditable agent behavior

Production Deployment, Monitoring, and Incident Response

Secure deployment patterns for single and multi-agent systems

Continuous monitoring for drift, anomalies, and abuse

Logging, audit trails, and forensic readiness for agent actions

AI security incident response playbooks and recovery procedures

Case studies of real-world AI breaches and lessons learned

Capstone and Synthesis

Reviewing the multi-modal multi-agent system built throughout the course

End-to-end pipeline review: design, build, secure, govern, deploy

Self-assessment of the system against NIST AI RMF functions

Forward outlook on emerging trends in agentic AI and AI security

Summary and Next Steps