Agentic AI Architectures¶

📖 NVIDIA Agentic AI Guide - Introduction to AI agent concepts and patterns

Agent Design Patterns¶

ReAct (Reasoning + Acting)¶

Core Loop: 1. Thought - Reason about the current state and determine next action 2. Action - Execute a tool or API call 3. Observation - Process the result 4. Repeat until task is complete or max iterations reached

Characteristics: - Most popular general-purpose agent pattern - Interleaves reasoning and action steps - Grounded in external tool feedback - Works well for open-ended tasks with tool access - Can get stuck in loops without proper termination conditions

When to Use: - General-purpose question answering with tools - Information retrieval and synthesis tasks - Tasks where the steps are not known in advance - Interactive debugging and exploration

Plan-and-Execute¶

Two Phases: 1. Planning phase - Generate a complete step-by-step plan 2. Execution phase - Execute each step sequentially, collecting results

Re-planning: - Monitor execution progress against the plan - Detect when results deviate from expectations - Generate a revised plan incorporating new information - Balance plan stability vs adaptation

When to Use: - Complex multi-step workflows with known structure - Tasks where order of operations matters - Situations requiring predictable execution paths - Long-running tasks that benefit from upfront planning

Reflexion¶

Loop: 1. Execute - Attempt the task 2. Evaluate - Check results against success criteria 3. Reflect - Identify what went wrong and how to improve 4. Retry - Use reflections to guide the next attempt 5. Store - Save reflections in memory for future reference

When to Use: - Tasks with clear success/failure criteria - Code generation and debugging - Mathematical problem solving - Tasks that benefit from iterative improvement

📖 NVIDIA Agent Blueprints - Reference implementations for agent patterns

Additional Patterns¶

Tool-Augmented Generation: - LLM decides when to invoke tools during generation - Supports parallel calls for independent operations - Sequential calls when results feed into subsequent decisions

Multi-Agent Debate: - Multiple agents independently generate responses - Agents critique and refine each other's outputs - Iterate until convergence or round limit - Improves accuracy on complex reasoning tasks

Memory Systems¶

Short-Term Memory (Working Memory)¶

• Current conversation history and recent messages
• Tool call results and observations
• Limited by the model's context window
• Managed through message history truncation or summarization

Long-Term Memory¶

• Persistent facts and learned information
• Implemented with vector databases for semantic retrieval
• Structured knowledge graphs for relationship tracking
• Survives across conversation sessions
• Requires explicit write and retrieval operations

Episodic Memory¶

• Records of past task executions and their outcomes
• Successful strategies indexed by task type
• Failed approaches to avoid repeating mistakes
• Enables learning from experience over time

Memory Management¶

• Summarization - Compress older messages to save context space
• RAG retrieval - Fetch relevant long-term memories on demand
• Priority retention - Keep high-importance information longer
• Consolidation - Periodically merge and deduplicate memories
• Eviction policies - Remove stale or low-value entries

📖 LangChain Memory - Memory management patterns for agents

State Management¶

Conversation State¶

• Track user intent and dialog progression across turns
• Maintain task context (what has been done, what remains)
• Handle interruptions and topic changes gracefully

Agent Internal State¶

• Current goals, sub-goals, and plans
• Intermediate results and partial computations
• Confidence levels and uncertainty tracking
• Tool availability and usage quotas

Multi-Agent Shared State¶

• Shared blackboard for inter-agent communication
• Task assignment and completion tracking
• Conflict resolution when agents disagree
• Synchronized access to shared resources

State Persistence¶

• Serialize state for long-running tasks
• Resume from checkpoints after failures
• Versioned state for rollback capabilities
• Session management for concurrent users

Architecture Selection Guide¶

Key Exam Concepts¶

• Know the differences between ReAct, Plan-and-Execute, and Reflexion
• Understand when each pattern is most appropriate
• Memory types - short-term, long-term, episodic and their implementations
• State management challenges in multi-turn and multi-agent settings
• Architecture trade-offs - predictability vs flexibility vs complexity