Agent = LLM + Context + Tools

agisoul is an independent AGI lab: a personal project about how LLM agents already make life, work, and small business easier. One author, no research institute: I’m breaking down the book “Deep Understanding of AI Agent” (Li Bojie) into atomic notes, assembling mini-agents, running ideas in practice; applied cases — including pitfalls — are collected in a sub-project of services.

Inside: 130+ atomic notes — one note = one idea, all linked in a single network; a mini-course “Agent from Scratch” and a journal-blog. How to navigate: sections and search — in the sidebar on the left, the entire network of connections — on the site map, Telegram channels — by the ✈ icon next to the sections.

📚 Site Sections

📖 Wiki — all 11 chapters of the book: list below, terms — in [Glossary]
🎓 [Learning] — mini-courses and step-by-step guides
📰 [Blog] — reviews and digests; applied cases — in a sub-project of services
👤 [About the project] — mission and contacts

🚀 Where to start

Three routes — for three audiences.

Just use — reviews without code: how models are arranged and what to adopt in everyday life — [Reviews]; unfamiliar terms — in 📖 [Glossary]
Learning to code — mini-course [“Agent from Scratch”]: 5 lessons, result — your own CLI agent
Building seriously — [Agent Formula] → [MCP Protocol] → [Three levels of the evaluation system]; the entire network of connections — on the map

🕒 Latest Updates

September 6, 2026 — applied cases highlighted in a sub-project of services (separate subdomain being prepared); reviews and digests remain in the blog
September 3, 2026 — focus on three audiences (no-code users, beginner programmers, small business): main page routes have been redesigned, cases have become the practical core of the blog, a case “anti-bot against scrapers” has been released (moved to the sub-project of services)
September 2, 2026 — positioning clarified: agisoul is an independent AGI lab, a personal project of one author; updated the main page, [About the project] (disclaimers added) and the [Blog] hub
August 25, 2026 — categories defined in the “Blog” and the first case “how agisoul counts time on the page” was released (cases later moved to the sub-project of services)
August 24, 2026 — the mini-course [“Agent from Scratch”] started in the “Learning” section: 5 lessons — theory with block diagrams and code in Python/JS, result — your own CLI agent
August 24, 2026 — “Learning”, “Blog” and “About the project” sections opened; added [Glossary] — 19 terms in one line

01 Agent Basics (Chapter 1)

Agent formula, ReAct cycle, Harness and fences — how the agent is arranged from the inside.

[Agent Formula]
[Observation and Action Spaces]
[Context as the Agent’s Eyes]
[Five Categories of Tools]
[Calling Tools]
[ReAct Cycle]
[Agent Cycle in Code]
[Three Levels of Ability Upgrade]
[Harness Engineering]
[Five Harness Functions]
[Evolution of Engineering Paradigms]
[Orchestration - Workflow and Autonomous Agent]
[Fences]
[ACI Design]

02 Context Engineering (Chapter 2)

What context consists of, KV Cache and Prompt Cache, chat templates and prompts.

[Context Engineering]
[Five Components of Context]
[Four Roles of API Messages]
KV Cache
Prompt Cache
[Three Rules of Friendliness to KV Cache]
Chat Template
[Prompt Engineering]
[Prompt Injection]
Agent Skills
[Agent Status String]
[Context Compression]

03 User Memory and Knowledge Base (Chapter 3)

Memory hierarchies, RAG, embeddings, hybrid search and GraphRAG.

[Three-Level Memory Evaluation]
[Memory Hierarchy]
[User Memory Formats]
User as Code
[Cognitive Types of Memory]
RAG
[Document Chunking]
[Dense Embedding]
[Sparse BM25 Search]
[Hybrid Search]
[Structured Indexing]
RAPTOR
GraphRAG
[Agent RAG]
[Context-Aware Search]

04 Tools (Chapter 4)

Classification and description of tools, MCP, Sidecar and human in the loop.

[Tool Classification]
[Specialized Tool vs Skill]
[Balance of Tool Granularity]
[The Art of Describing Tools]
[Reliability of Parameter Transfer]
[MCP Protocol]
[MCP Security Risks]
[Sidecar Mechanism]
[Proposer-Reviewer]
[Human in the Loop]
[Event-Driven Asynchronous Agent]
[Interruptions and Placeholders]
[Proactive Tool Discovery]
[Dynamic Tool Loading]

05 Coding Agent and Code Generation (Chapter 5)

File system as the center of the agent, code as ability and as limitation.

[Seven Basic Tools]
[File System as the Center of the Agent]
[File System Paradigm for Knowledge]
Sessionless
[Code as a Metacapability]
[Code as a Thinking Tool]
[Code as a Limitation of Business Rules]
[Generative UI]
[Agent Self-Loading]
[Deadly Triad]
[Agent Loyalty]
[Semantic Command Parsing]
[Failures and Recovery]

06 Agent Evaluation (Chapter 6)

Metrics, LLM-as-a-Judge, cost and observability of agents.

[Three Levels of the Evaluation System]
[Evaluation Environments]
[Agent Evaluation Metrics]
LLM-as-a-Judge
[Four Principles of Rubric]
[Statistical Significance of Evaluation]
[Agent Cost Analysis]
[From Benchmark to Improvements]
[Agent Observability]
[Simulation Environment]

07 Model Post-Training (Chapter 7)

SFT, RLHF, GRPO, RLVR and distillation.

[Three Stages of Post-Training]
[SFT Memorizes - RL Generalizes]
[First SFT then RL]
[Data and Environment are More Important than Algorithm]
RLHF
[KL - mode-seeking vs mass-covering]
[RLVR and Verifiable Rewards]
[Reward for Process and Reward for Result]
GRPO
On-Policy Distillation
[OPSD Self-Distillation]

08 Continuous Agent Evolution (Chapter 8)

Learning from trajectories, update loops, safety of self-modification.

[Continuous Agent Evolution]
[Learning Signals from Trajectories]
[Four Carriers of Update]
[From Experience to Knowledge]
[Update Prompt and Skill]
[[From Experience to