Agentic AI

Agentic AI is the frontier where language models stop being question-answering tools and start being autonomous systems that plan, use tools, and complete multi-step tasks. These notes document the architecture, patterns, and engineering practices behind building reliable agentic systems.


Core Topics

Large Language Model Fundamentals

Transformer architecture, tokenization, context windows, attention mechanisms, fine-tuning vs. prompting, inference parameters.

Prompt Engineering

System prompts, few-shot examples, chain-of-thought, structured outputs, XML formatting, prompt injection and defense.

Tool Use & Function Calling

How LLMs call external tools, structured function schemas, tool result handling, error recovery patterns.

Agentic Architectures

ReAct (Reasoning + Acting), Plan-and-Execute, Reflexion, and other agent frameworks. When each pattern applies.

Multi-Agent Systems

Orchestrator-subagent patterns, agent handoffs, shared memory, agent communication, parallelism. Frameworks: LangGraph, AutoGen, Claude MCP.

Model Context Protocol (MCP)

MCP architecture, server/client model, tool definitions, resource exposure, sampling. Building and consuming MCP servers.

Memory Systems

In-context memory, external memory (vector databases, knowledge graphs), episodic vs. semantic memory, retrieval-augmented generation (RAG).

RAG Systems

Chunking strategies, embedding models, vector databases (Pinecone, Chroma, pgvector), retrieval strategies, re-ranking, hybrid search.

Evaluation & Reliability

Benchmarking agents, failure modes, hallucination mitigation, human-in-the-loop patterns, tracing and observability (LangSmith, Arize).

Agentic AI in Engineering

Agents for code generation, document analysis, technical report writing, grid data analysis, standards lookup, and engineering workflow automation.


Key Questions These Notes Answer

  • What is the difference between a chain and an agent?
  • How do I design a multi-agent system that doesn’t fail silently?
  • How does RAG work and when should I use it vs. fine-tuning?
  • How do I build a reliable agent that uses external tools?
  • What is MCP and how does it change how agents access capabilities?
  • How do I evaluate whether my agent is actually working correctly?

Prerequisites

Connects To

0 items under this folder.