Agentic AI Notebook
0What is an AI Agent?1Why LLMs Need Agents2Anatomy of an Agent3Agent Lifecycle4Core Concepts5Agent Capabilities6Types of Agents7Agent Architectures8Agent Terminology9Current Agent Landscape10Planning11Reflection12Multi Tool13Self Correction14Build First AI Agent
Agent Foundations
Phase 4Module 5 of 15

Core Concepts

Shared vocabulary lets you evaluate any framework critically instead of following hype.

Core concepts are chess pieces — once you know how each moves, you can play on any board.

Visual Workflows

Start here — scroll inside each diagram frame to explore, then use + / − to zoom up to 200% if needed.

100%
Loading diagram...

Scroll inside the frame to explore · use + / − to zoom up to 200%

Perception-Action Loop

100%
Loading diagram...

Scroll inside the frame to explore · use + / − to zoom up to 200%

The fundamental cycle every agent architecture implements.

Key Takeaways

  • 1.Perception-action: observe think act repeat.
  • 2.Planning: ReAct interleaved or plan-and-execute upfront.
  • 3.Memory: working episodic semantic tiers.
  • 4.Multi-agent: supervisor delegates to workers.
  • 5.HITL: human approval on high-stakes actions.

Real Example

Scenario

EU AI Act research brief: supervisor routes to a web-scraper worker, then a summarizer, then pauses at HITL before publishing to Notion.

What you would do

Perception-action: scraper observes URLs, supervisor decides the next worker. Plan-and-execute: upfront outline (search → summarize → cite → publish). HITL blocks the Notion write until legal reviews. Monitor: routing accuracy and unauthorized publish attempts.

Practice Task

For the EU AI Act brief, label each step as ReAct, plan-and-execute, supervisor routing, or HITL. Write one sentence on why HITL sits where it does.

Code Walkthrough

Highlighted lines show where Core Concepts happens in the code.

Core Concepts
1SYSTEM = "You are a ReAct agent. Reply with Thought, then Action, then wait for Observation."2
3messages = [4    {"role": "system", "content": SYSTEM},5    {"role": "user", "content": "Find 3 official EU AI Act sources and summarize obligations"},6]7
8# One ReAct iteration9thought_action = llm(messages)  # "Thought: search official sources\nAction: web_search({'q': 'EU AI Act official'})"10observation = web_search("EU AI Act site:europa.eu")11messages.append({"role": "assistant", "content": thought_action})12messages.append({"role": "user", "content": f"Observation: {observation}"})

Cheat Sheet

Quick recap

quick ref
  • •ReAct = think + act each step
  • •Plan-execute = plan upfront
  • •Memory tiers serve different horizons
  • •Supervisor = router to specialists
  • •HITL for destructive actions

Common Mistakes

  • ✕Skipping evaluation for Core Concepts before production
  • ✕No logging or tracing around core concepts steps
  • ✕Ignoring cost and latency implications