COMPREHENSIVE SYLLABUS

Claude AI & Agentic AI

From Foundations to Autonomous Agent Systems

Level	Beginner to Advanced
Format	Lectures, Labs, Projects, Capstone
Prerequisites	Basic Python; Familiarity with APIs
Certification	Certificate of Completion upon graduation

 

 

1. Course Overview

This syllabus provides a complete educational pathway through Claude AI—Anthropic’s large language model family—and Agentic AI, covering autonomous agent architectures, multi-agent systems, and real-world deployments. Students progress from foundational theory to building production-grade AI agent pipelines.

 

1.1  Course Objectives

• Understand the architecture, training philosophy, and capabilities of Claude AI models
• Master prompt engineering, Constitutional AI principles, and safety-first design
• Build agentic AI systems using Claude as a reasoning core
• Design, implement, and evaluate multi-agent pipelines with tools and memory
• Apply agentic AI to real-world use cases: research, coding, customer service, automation
• Navigate ethical, legal, and safety considerations in autonomous AI deployment

 

1.2  Learning Outcomes

1. Explain how Claude models work, including RLHF and Constitutional AI
2. Write effective prompts and system instructions for diverse tasks
3. Build Claude-powered applications using the Anthropic API
4. Design and deploy autonomous agents with tools, memory, and planning
5. Architect multi-agent systems and evaluate their reliability
6. Apply responsible AI practices to agentic deployments

 

 

 

Module 1: Introduction to Claude AI & Large Language Models

 

 

 

Learning Goals

• Understand the landscape of modern LLMs and where Claude fits
• Explain transformer architecture at a conceptual level
• Describe Anthropic’s mission and safety-first approach

 

Topics

#	Topic	Description
1.1	History of AI & LLMs	From rule-based systems to GPT, PaLM, and Claude — the evolution of language models
1.2	Transformer Architecture	Attention mechanisms, tokenisation, embeddings, encoder-decoder models explained accessibly
1.3	Claude Model Family	Haiku, Sonnet, Opus — capability tiers, context windows, and use-case fit
1.4	Anthropic & AI Safety	Company mission, responsible scaling policy, interpretability research
1.5	Constitutional AI (CAI)	How Claude is trained with principles; RLHF vs. RLAIF; harmlessness + helpfulness tension
1.6	Claude’s Personality & Values	Honesty, epistemic humility, refusal behaviour, and what makes Claude different

 

Lab 1 — Explore the Claude API

• Set up Anthropic API keys and Python SDK
• Send your first messages; inspect token counts and latency
• Compare Haiku vs. Sonnet vs. Opus on identical prompts

 

Module 2: Prompt Engineering & System Design

 

 

 

Learning Goals

• Write high-quality, reliable prompts for diverse tasks
• Design system prompts for specialised Claude deployments
• Apply advanced techniques: chain-of-thought, few-shot, XML structuring

 

Topics

#	Topic	Description
2.1	Prompt Fundamentals	Anatomy of a prompt: role, context, task, format, constraints
2.2	System Prompts	Persona design, persistent instructions, operator vs. user roles
2.3	Few-shot Prompting	Providing examples effectively; when few-shot helps vs. hurts
2.4	Chain-of-Thought (CoT)	Step-by-step reasoning; zero-shot CoT; extended thinking with Claude
2.5	XML & Structured Output	Asking Claude to return JSON, XML, tables; parsing strategies
2.6	Prompt Chaining	Breaking tasks into sub-tasks; passing context between calls
2.7	Prompt Injection & Safety	Adversarial prompts, jailbreak patterns, and defensive design
2.8	Evaluation & Iteration	Systematic prompt testing; A/B evaluation; evals with LLM-as-judge

 

Lab 2 — Build a Prompt Library

• Create a reusable prompt template library for 5 use cases
• Implement CoT for a multi-step reasoning task
• Build an LLM-as-judge evaluator using Claude

 

Module 3: Claude API & Application Development

 

 

 

Learning Goals

• Integrate Claude into Python and web applications
• Handle streaming, tokens, and cost management
• Build production-ready wrappers with error handling and retries

 

Topics

#	Topic	Description
3.1	Anthropic SDK Deep Dive	Python & TypeScript SDK; messages API; model parameters
3.2	Streaming Responses	Server-sent events; real-time output; chunked delivery
3.3	Token Management	Counting tokens; context window limits; truncation strategies
3.4	Cost Optimisation	Caching, prompt compression, model routing, batching
3.5	Vision & Multimodal	Sending images; document analysis; reading charts and screenshots
3.6	Tool Use (Function Calling)	Defining tools, receiving tool calls, returning results, multi-turn tool loops
3.7	Error Handling & Retries	Rate limits, timeouts, exponential back-off, graceful degradation
3.8	Production Patterns	Logging, monitoring, secrets management, API gateway setup

 

Lab 3 — Claude-Powered Document Analyser

• Build a document Q&A app with PDF + image support
• Implement streaming output with a simple web frontend
• Add token tracking and cost dashboard

 

Module 4: Safety, Ethics & Responsible AI with Claude

 

 

Learning Goals

• Understand Claude’s safety mechanisms and their limits
• Apply responsible deployment practices
• Navigate real-world ethical dilemmas in AI product design

 

Topics

#	Topic	Description
4.1	Harm Avoidance Framework	Hardcoded vs. softcoded limits; operator vs. user trust levels
4.2	Honesty & Deception	Non-deception, calibration, transparency, avoiding sycophancy
4.3	Bias & Fairness	Sources of model bias; mitigation at training vs. inference time
4.4	Content Moderation	Building moderation pipelines with Claude; human-in-the-loop
4.5	Privacy & Data Handling	PII risks, data minimisation, GDPR/CCPA considerations
4.6	Legal & Compliance	Copyright, liability, disclosure requirements for AI-generated content
4.7	Dual-Use Risks	When capabilities become dangers; red-teaming your own system
4.8	AI Governance	Policy landscape, EU AI Act, NIST AI RMF, internal governance

 

Midterm Project — Claude-Powered Application

Build a complete Claude-powered application with a real use case. Must include: system prompt design, multi-turn conversation, at least one tool, safety guardrails, and a reflection on ethical considerations. Presentations in Week 8.

 

 

PART B: AGENTIC AI — Autonomous Systems & Pipelines

 

Module 5: Foundations of Agentic AI

 

 

Learning Goals

• Define agentic AI and distinguish it from standard LLM use
• Understand core agent loop: perceive → plan → act → observe
• Survey major agent architectures and frameworks

 

Topics

#	Topic	Description
5.1	What is Agentic AI?	Definition; autonomy spectrum; task complexity vs. oversight tradeoffs
5.2	The ReAct Pattern	Reason + Act interleaved; trace analysis; when ReAct succeeds/fails
5.3	Agent Loop Architecture	Environment, observation, thought, action, feedback loop in detail
5.4	Tool & Action Spaces	Search, code execution, file I/O, web browsing, API calls as agent tools
5.5	Memory Types	In-context, episodic (vector DBs), semantic, procedural memory stores
5.6	Planning Strategies	Chain-of-thought, tree-of-thought, goal decomposition, MCTS
5.7	Agent Frameworks Overview	LangChain, LlamaIndex, AutoGen, CrewAI, Claude’s native tool use
5.8	Benchmarks & Evaluation	WebArena, AgentBench, GAIA; measuring agent reliability

 

Lab 4 — Build a ReAct Agent

• Implement a ReAct agent using Claude + Python from scratch
• Give it web search and calculator tools
• Trace and debug agent reasoning steps

 

Module 6: Memory, Knowledge & Retrieval for Agents

 

 

Topics

#	Topic	Description
6.1	RAG Architecture	Retrieval-Augmented Generation: chunking, embedding, retrieval, reranking
6.2	Vector Databases	Pinecone, Weaviate, Chroma, pgvector — setup and querying
6.3	Embedding Models	OpenAI, Cohere, sentence-transformers; choosing the right model
6.4	Hybrid Search	Combining dense + sparse retrieval; BM25 + vector fusion
6.5	Agent Working Memory	Sliding window, summarisation, context distillation strategies
6.6	Long-term Memory	Storing and retrieving agent experiences; episodic memory design
6.7	Knowledge Graphs	Integrating structured knowledge; Neo4j + LLM reasoning
6.8	Memory Consistency & Privacy	Avoiding hallucination from stale memory; PII in memory stores

 

Lab 5 — RAG-Powered Research Agent

• Build a document ingestion pipeline (PDF → chunks → embeddings)
• Connect Claude to a vector DB for grounded Q&A
• Implement context distillation for long research sessions

 

Module 7: Multi-Agent Systems & Orchestration

 

 

 

Topics

#	Topic	Description
7.1	Multi-Agent Motivation	Why single agents fail at complex tasks; division of labour
7.2	Orchestrator–Subagent Pattern	Claude as orchestrator; delegating to specialised subagents
7.3	Agent Communication	Message passing, shared state, blackboard architectures
7.4	CrewAI & AutoGen	Framework-level multi-agent: roles, crews, conversations
7.5	Model Context Protocol (MCP)	Anthropic’s MCP standard; building MCP servers and clients
7.6	Parallelism & Fan-Out	Parallel subagent execution; result aggregation; race conditions
7.7	Conflict Resolution	Disagreement between agents; voting, consensus, hierarchy
7.8	Reliability & Error Recovery	Checkpointing, retry logic, fallback agents, dead-letter queues

 

Lab 6 — Multi-Agent Research Pipeline

• Build a 3-agent system: Researcher, Writer, Editor using Claude
• Implement MCP server for tool sharing between agents
• Add orchestrator with task routing and error recovery

 

Research Paper Due — Week 12

15-page paper on a chosen agentic AI topic. Topics may include: reliability of autonomous agents, memory architecture trade-offs, MCP ecosystem analysis, multi-agent coordination protocols, or agentic AI in a specific domain (healthcare, legal, finance).

 

Module 8: Agentic AI in Production

 

 

Topics

#	Topic	Description
8.1	Production Architecture	Microservices vs. monolith for agents; queue-based workflows
8.2	Observability & Tracing	LangSmith, Langfuse, Arize; distributed tracing for agent loops
8.3	Latency & Cost at Scale	Async execution, caching, model routing, cost budgets per task
8.4	Human-in-the-Loop (HITL)	Interrupts, approvals, escalation triggers; keeping humans informed
8.5	Security for Agents	Prompt injection in agentic contexts; sandboxing code execution; secrets
8.6	Agent Testing Strategies	Unit tests for tools; integration tests for loops; chaos engineering
8.7	Deployment Patterns	Docker, Kubernetes, serverless functions for agents; long-running jobs
8.8	Agent Versioning & CI/CD	Prompt versioning, model pinning, regression suites, canary deploys

 

Lab 7 — Production-Ready Agent Deployment

• Containerise a Claude agent with Docker
• Add distributed tracing with Langfuse
• Implement HITL approval step for high-stakes actions

 

Module 9: Advanced Topics & Future Directions

 

 

 

Topics

#	Topic	Description
9.1	Frontier Models & Reasoning	o1/o3-style extended thinking; Claude’s reasoning modes; what’s next
9.2	Multimodal Agents	Agents that see, hear, and generate images/video; computer-use agents
9.3	Embodied & Robotic AI	Connecting agentic AI to physical systems; ROS + LLM integration
9.4	AI-Assisted Software Engineering	Claude Code, Devin-style agents; the future of developer tools
9.5	Autonomous Research Agents	AI scientists; hypothesis generation; lab automation
9.6	Economic & Social Impact	Labour displacement, new job categories, AI as colleague
9.7	Long-Horizon Safety	Alignment for agentic systems; corrigibility; value lock-in risks
9.8	Open Problems & Research Frontiers	Reliability, world models, genuine autonomy — what remains unsolved