Course Outline
Day 1
Anatomy of a Modern AI Agent
Beyond chatbots, agents are autonomous systems capable of reasoning and acting.
Paradigms include reactive, proactive, hybrid, and goal-directed agents.
Core components encompass perception, planning, memory, tool use, and action.
Single-agent versus multi-agent design tradeoffs are explored for government applications.
Agent Frameworks and the Modern Stack
Frameworks such as LangChain, LlamaIndex, AutoGen, and CrewAI are compared with classical frameworks like JADE and SPADE.
Choosing a framework based on production requirements for government use is discussed.
Tool Calling, Function Calling, and Structured Outputs
Hands-on session: scaffolding a single Python agent with tool calls.
Multi-Agent System Architectures
Designs for centralized, decentralized, hybrid, and layered multi-agent systems (MAS) are examined.
Communication protocols such as FIPA ACL and modern equivalents are covered.
Coordination patterns, including planning, negotiation, and synchronization, are discussed.
Emergent behavior and self-organization in agent populations are explored.
Decision-Making and Learning in Agents
Game theory for cooperative and competitive interactions among agents is introduced.
Reinforcement learning in multi-agent environments is covered.
Transfer learning and knowledge sharing across agents are discussed.
Conflict resolution and trust between coordinating agents are addressed.
Day 2
Multi-Modal Foundations for Agents
Multi-modal AI as a unified workflow across text, image, speech, and video is explored.
Leading multi-modal models such as GPT-4 Vision, Gemini, Claude, and Whisper are discussed.
Techniques for combining modalities within an agent's reasoning loop are covered.
Tradeoffs in latency, cost, and accuracy in multi-modal pipelines are examined.
Building the Perception Layer
Image processing techniques for classification, captioning, and object detection are introduced.
Speech recognition using Whisper ASR and streaming transcription is covered.
Text-to-speech synthesis and natural voice interaction are discussed.
Connecting perception outputs to LLM-driven reasoning and tool selection is explored.
Hands-On - Building a Multi-Modal Agent in Python
Defining the agent's task, context window, and tool inventory is covered.
Wiring up GPT-4 Vision and Whisper APIs end-to-end is demonstrated.
Implementing memory, state, and conversation management is addressed.
Safely adding tool calls that produce real-world side effects is discussed.
Hands-On - Orchestrating a Multi-Agent System
Composing specialized agents with AutoGen or CrewAI is demonstrated.
Defining roles, responsibilities, and inter-agent communication protocols is covered.
Resource allocation and coordination in a simulated environment are addressed.
Logging agent reasoning, tool calls, and decisions for inspection and audit is discussed.
Day 3
Threat Surface of Production AI Agents
Unique vulnerabilities of agentic AI compared to traditional software are explored.
Attack surfaces at data, model, prompt, tool, output, and interface layers are discussed.
Threat modeling for agent-based systems with autonomous tool use is covered.
Comparing AI cybersecurity practices to traditional cybersecurity is addressed.
Adversarial Attacks Hands-On
Methods such as FGSM, PGD, and DeepFool for adversarial examples are introduced.
White-box versus black-box attack scenarios are covered.
Model inversion and membership inference attacks are discussed.
Data poisoning and backdoor injection during training are addressed.
Prompt injection, jailbreaking, and tool misuse in LLM-based agents are explored.
Defensive Techniques and Model Hardening
Adversarial training and data augmentation strategies are introduced.
Defensive distillation and other robustness techniques are covered.
Input preprocessing, gradient masking, and regularization are discussed.
Differential privacy, noise injection, and privacy budgets are addressed.
Federated learning and secure aggregation for distributed training are explored.
Hands-On with the Adversarial Robustness Toolbox
Simulating attacks against the multi-modal agent built on Day 2 is demonstrated.
Measuring robustness under perturbation and quantifying degradation is covered.
Applying defenses iteratively and re-evaluating attack success rates is addressed.
Stress-testing tool-call pathways and prompt injection vectors is explored.
Day 4
Risk Management Frameworks for AI
The NIST AI Risk Management Framework (RMF) functions of govern, map, measure, and manage are introduced.
ISO/IEC 42001 and emerging AI-specific standards are discussed.
Mapping AI risk to existing enterprise GRC frameworks is covered.
AI accountability, auditability, and documentation requirements for government are addressed.
Regulatory Compliance for Agentic Systems
The EU AI Act's risk tiers, prohibited uses, and obligations for high-risk systems are introduced.
GDPR and CCPA implications for agent data pipelines are covered.
The U.S. Executive Order on Safe, Secure, and Trustworthy AI is addressed.
Sector-specific guidance for finance, healthcare, and public services is explored.
Third-party risk and supplier AI tool usage are discussed.
Ethics, Bias, and Explainability
Bias detection and mitigation across agent perception and reasoning are introduced.
Explainability and transparency as security-relevant properties are covered.
Fairness, downstream harm, and responsible deployment are addressed.
Designing inclusive, auditable agent behavior is explored.
Production Deployment, Monitoring, and Incident Response
Secure deployment patterns for single and multi-agent systems are introduced.
Continuous monitoring for drift, anomalies, and abuse is covered.
Logging, audit trails, and forensic readiness for agent actions are addressed.
AI security incident response playbooks and recovery are discussed.
Case studies of real-world AI breaches and lessons learned are explored.
Capstone and Synthesis
Reviewing the multi-modal multi-agent system built across the course is covered.
An end-to-end pipeline review, from design to deployment, is conducted.
Self-assessment of the system against NIST AI RMF functions is addressed.
A forward outlook on emerging trends in agentic AI and AI security is explored.
Summary and Next Steps
Requirements
Targeted Audience
This guidance is intended for AI engineers and architects developing agentic systems for production use in government and other sectors. It also targets cybersecurity, risk, and compliance professionals responsible for ensuring AI assurance in highly regulated industries such as finance, healthcare, and consulting. Additionally, senior developers and solution leads who are integrating multi-modal and multi-agent capabilities into enterprise platforms will find this information valuable.
Testimonials (3)
The trainer is patient and very helpful. He knows the topic well.
CLIFFORD TABARES - Universal Leaf Philippines, Inc.
Course - Agentic AI for Business Automation: Use Cases & Integration
Good mixvof knowledge and practice
Ion Mironescu - Facultatea S.A.I.A.P.M.
Course - Agentic AI for Enterprise Applications
The mix of theory and practice and of high level and low level perspectives
