Quick Reference Cards
Printable one-page summaries for each week. Use Ctrl+P to print individual sections.
Week 1: Introduction to Agentic AI
Key Concepts
- Agent = perceive + reason + act
- ReAct = Reasoning + Acting
- Trajectory = (s, a, o) sequence
ReAct Loop
- Thought: reasoning step
- Action: tool/API call
- Observation: result
Key Papers
- ReAct (Yao et al., 2023)
- Agent Survey (Wang, 2024)
Code Pattern
- while not done:
- thought = think(obs)
- action = act(thought)
- obs = env.step(action)
Week 2: LLM Foundations
Prompting Strategies
- Zero-shot: no examples
- Few-shot: with examples
- CoT: step-by-step
Chain-of-Thought
- "Let's think step by step"
- Emergent in 100B+ models
- Best for math/logic
Advanced Methods
- Self-Consistency: majority vote
- ToT: tree exploration
- Self-Ask: sub-questions
Trade-offs
- CoT: +accuracy, +tokens
- SC: +robust, +cost
- ToT: +complex, +latency
Week 3: Tool Use
Function Calling
- Define JSON schema
- Model generates call
- Execute and return
MCP Protocol
- Tools: functions to call
- Resources: data to read
- Prompts: templates
Tool Design
- Clear, specific names
- Typed parameters
- Detailed descriptions
Key Papers
- Toolformer (Schick, 2023)
- Gorilla (Patil, 2023)
Week 4: Planning & Reasoning
Reflexion
- Actor: generates actions
- Evaluator: success signal
- Self-Reflect: verbal feedback
Memory Types
- Short-term: current task
- Long-term: persistent
- Episodic: experiences
LATS
- Tree search over actions
- MCTS for exploration
- Best path selection
Plan-and-Solve
- 1. Decompose problem
- 2. Plan subtasks
- 3. Execute each
Week 5: Multi-Agent Systems
Topologies
- Hub-and-spoke: central
- Mesh: peer-to-peer
- Hierarchical: layers
Agent Roles
- Orchestrator: coordinates
- Specialist: domain expert
- Critic: reviews output
Communication
- Message passing
- Shared blackboard
- Structured protocols
Frameworks
- AutoGen: conversational
- MetaGPT: structured
- CrewAI: role-based
Week 6: Agent Frameworks
LangGraph
- StateGraph: state machine
- Nodes: functions
- Edges: transitions
- Checkpointing: save state
AutoGen
- ConversableAgent
- GroupChat patterns
- Code execution
CrewAI
- Role: agent identity
- Goal: objective
- Backstory: context
Selection Criteria
- State needs?
- Multi-agent?
- Production ready?
Week 7: Advanced RAG
Self-RAG
- Reflection tokens
- Retrieve: yes/no
- Relevance: check
- Support: verify
CRAG
- Correct: use docs
- Incorrect: web search
- Ambiguous: refine
RAPTOR
- Hierarchical summaries
- Tree structure
- Multi-level retrieval
Query Enhancement
- Decomposition
- Expansion
- Rewriting
Week 8: GraphRAG
Components
- Entity extraction
- Relationship detection
- Graph construction
Search Types
- Local: specific entities
- Global: community themes
Community Detection
- Cluster entities
- Summarize clusters
- Hierarchical index
HippoRAG
- Hippocampus-inspired
- Pattern separation
- Associative retrieval
Week 9: Hallucination Prevention
Hallucination Types
- Intrinsic: contradicts input
- Extrinsic: unsupported
- Factual: wrong facts
CoVe Pipeline
- 1. Generate response
- 2. Generate questions
- 3. Answer independently
- 4. Revise if needed
FActScore
- Decompose to atoms
- Verify each claim
- Calculate precision
Mitigation
- Retrieval grounding
- Confidence calibration
- Self-consistency
Week 10: Agent Evaluation
Benchmarks
- AgentBench: 8 envs
- SWE-bench: GitHub
- WebArena: web tasks
- GAIA: general AI
Metrics
- Success rate
- Pass@k
- Task completion %
LLM-as-Judge
- Automated scoring
- Pairwise comparison
- Rubric-based
Challenges
- Contamination
- Generalization
- Cost
Week 11: Domain Applications
Code Agents
- Devin: autonomous
- AlphaCodium: flow
- Sandboxing critical
Finance Agents
- SEC compliance
- Audit logging
- Risk controls
Healthcare Agents
- HIPAA compliance
- Evidence-based
- Human oversight
Safety Patterns
- Sandboxing
- Rate limiting
- Approval gates
Week 12: Research Frontiers
Generative Agents
- Memory stream
- Reflection
- Planning
- Emergent behavior
Voyager
- Skill library
- Curriculum learning
- Open-ended exploration
Open Problems
- Long-horizon planning
- Grounded world models
- Safe autonomy
Constitutional AI
- Principle-based
- Self-improvement
- AI feedback