Phase 1 Β· Cognitive Foundation

The Agent Mind Map

Before writing the first line of Agent code, build a complete cognitive framework. This includes LLM principles, Agent paradigms, key concepts, plus a readable, reproducible, thinkable knowledge system.

Knowledge Panorama

The learner's mental model panorama

See the connections between knowledge nodes at a glance. Choose where to dive based on your progress.

graph TD
  Start([πŸ‘€ Start here]) --> A[πŸ€– What is an LLM?]
  A --> B[πŸ“ Prompt Engineering]
  A --> C[πŸ”¬ Transformer & Attention]
  B --> D[πŸ› οΈ Function Calling]
  D --> E[πŸ€– Agent Paradigms]
  C --> F[🧠 Reasoning]
  F --> E
  E --> G[🌊 ReAct Loop]
  E --> H[πŸ“‹ Plan-and-Execute]
  E --> I[πŸ‘₯ Multi-Agent]
  G --> J[πŸ’Ύ Memory Systems]
  H --> J
  I --> J
  J --> K[🧰 Tools & Orchestration]
  K --> L[πŸ“š RAG]
  L --> M[🌐 Project Blueprint]
  K --> M
  I --> M
  M --> N([🎯 Deliver Large Agent])

  classDef critical fill:#7c5cff,stroke:#00d4ff,color:#fff,stroke-width:2px;
  classDef important fill:#00d4ff,stroke:#00ffa3,color:#000;
  classDef useful fill:#00ffa3,stroke:#7c5cff,color:#000;
  classDef final fill:#ff6ec7,stroke:#ffb84d,color:#fff,stroke-width:3px;

  class A,B,D,E critical;
  class G,H,I,J important;
  class C,F,K,L useful;
  class M,N final;
      
Core Concepts

Understanding the basic elements of an Agent

From the LLM "brain", to tools, memory, and planning β€” dissect each key Agent capability.

Foundation 01

Large Language Models (LLMs)

The Agent's "brain". Based on the Transformer architecture, learns statistical patterns from massive text, generates the next token autoregressively.

  • Evolution from GPT-3 to GPT-4o / Claude / Gemini
  • Tokens, context window, temperature, Top-p
  • Hallucinations and boundaries: what LLMs can't do
Foundation 02

Prompt Engineering

Steering LLM output through carefully constructed instructions. The "art" of talking to LLMs.

  • Zero-shot / Few-shot / Chain-of-Thought
  • System Prompt role and best practices
  • Structured Output and JSON mode
Foundation 03

Function Calling

Lets LLMs break beyond pure text, calling external APIs, querying databases, executing code. The source of Agent "action capability".

  • OpenAI Function Calling / Tool Use
  • Tool description writing specifications
  • Multi-tool selection and parameter generation
Core 04

Reasoning

More than pattern matching. LLMs "reason" between tokens via CoT, Self-Consistency, Tree-of-Thoughts.

  • CoT / ReAct / Self-Refine
  • Inference-time Compute
  • o1/o3 reasoning model paradigm shift
Core 05

Memory

Short-term conversation buffer, long-term fact storage, vector semantic memory β€” key to Agent continuous evolution.

  • Short-term: in-context conversation
  • Long-term: vector DB + knowledge graph
  • Episodic, semantic, procedural memory
Core 06

Planning

Decompose complex tasks into executable steps, choose appropriate tools, handle exceptions during execution and dynamically adjust plans.

  • Plan-and-Execute mode
  • Task Decomposition
  • Reflection and Re-planning
Architecture Paradigms

Mainstream Agent running paradigms

Understand each paradigm's design motivation, use cases, and typical implementations.

ReAct Loop
Plan-and-Execute
Multi-Agent
Graph (LangGraph)

ReAct: Think - Act - Observe Loop

The most classic Agent paradigm. Each loop: the model first thinks (Thought) about current state, decides the next action, executes, observes the result, and enters the next loop until the task is complete.

Thought: Reasoning

LLM analyzes current state, decides what to do next. This determines the Agent's "intelligence".

Action: Acting

Choose tools, generate parameters, call external systems. The Agent's "action capability".

Observation

Receive tool return results as input for the next loop.

Loop or Terminate

If task is complete, output the final answer; otherwise return to Thought to continue.

i
Key Insight

ReAct's "thought quality" directly determines Agent upper limit. This is why OpenAI o1/o3 reasoning models continue to lead in large tool-call scenarios.

Plan-and-Execute: Plan first, then execute

Suitable for complex, decomposable tasks. First use LLM to generate a complete execution plan, then execute step by step. Typical representatives: BabyAGI, AutoGPT, LangGraph planner.

🎯 Planner
Decompose tasks, generate steps
βš™οΈ Executor
Execute plan steps one by one
πŸͺž Replanner
Adjust plan based on execution feedback
  • More suitable for complex tasks requiring long chains of reasoning
  • Cached plan results save tokens
  • Easier to debug: plan is readable
  • Failure rollback is more controllable

Multi-Agent Collaboration: Divide and Conquer

Hand complex tasks to multiple specialized Agents who collaborate, debate, supervise. CrewAI, AutoGen, MetaGPT are representatives.

πŸ‘” Manager Agent
Assign tasks, supervise progress
πŸ§‘β€πŸ’» Engineer Agent
Focus on implementing specific tasks
🧐 Reviewer Agent
Quality checks and feedback
!
Trade-off Warning

Multi-Agent isn't always better. It increases token costs and complexity. Single Agent is more efficient when tasks are clear.

Graph Orchestration: Stateful Agent composition

LangGraph abstracts Agents as directed graphs (DAG) with nodes as actions and edges as conditions. Naturally supports state management, human intervention, parallel branches.

Core Advantages

  • Explicit state: Each node shares global state, can be checkpointed
  • Conditional branches: Determine next step based on state
  • Native loops: Graph structure allows nodes to return to themselves
  • Human-in-the-loop: Can pause and wait for human input
Self-Check

Have you truly mastered it?

Complete the checklist below. If anything is uncertain, return to the relevant section.

🧠 Phase 1 Self-Check List

  • βœ“
    Can explain what an AI Agent is to a non-technical person in one paragraph
    Not "an LLM that calls tools", but an autonomous system with perception, reasoning, action, and memory.
  • βœ“
    Can distinguish LLM, Agent, AI Assistant, AI Workflow differences
    The core of an Agent is "autonomous decision-making on what to do", while LLM is just "a model that's called".
  • βœ“
    Can describe ReAct, Plan-and-Execute, Multi-Agent paradigm differences and use cases
    Choose the most appropriate paradigm for specific tasks with reasons.
  • βœ“
    Can independently design a high-quality Tool description (with parameter schema)
    Tool description is key input for LLM decisions, treat it like API documentation.
  • βœ“
    Understand Function Calling's internal flow (user input to tool return)
    Not just "calling an API", but how the model parses, generates parameters, and routes back to context.
  • βœ“
    Know the essential difference between short-term and long-term memory and how to implement
    Short-term memory = context window content; long-term memory = persistent vector/structured data.

List complete? Move on to the next phase:

Phase 2 Β· Skills & Tools β†’

Foundation is solid

In Phase 2 you'll learn mainstream Agent frameworks, master tools and orchestration capabilities.

View Complete Learning Path