Reading Matrix — 3 Paths Through 46 Discussions (Builder / Architect / Leader)

[xg-gh-25]

---

46 articles, 5 themes, 3 reader paths. Don't read everything. Pick your path.

Base URL: https://github.com/xg-gh-25/SwarmAI/discussions/

Start Here (All Paths)

Before diving into any path below, read the Design Philosophy. It explains the "why" behind everything else — in ~10 minutes you'll understand the six pillars that drive every architectural choice.

Article	What you'll get	~Min
Design Philosophy — Six Pillars (EN)	The 6 core beliefs, how they earned from failures, why they compound	10
设计哲学 — 六根支柱 (CN)	同上中文版	10

Then pick your path ↓

---

Three Reader Paths

Path	You are...	Start here	Time
Builder	Using Claude Code / Cursor / Kiro daily, want to level up	Path A (bottom-up)	~45 min
Architect	Designing agent systems, evaluating frameworks	Path B (top-down)	~60 min
Leader	Making buy/build/invest decisions about AI tools	Path C (strategic)	~30 min

---

Path A: Builder

"I use agents daily, how do I get more from them?"

Step	Article	Why read this	~Min
1	What Is an Agent Harness? (Autonomy Levels L1-L5)	"Oh — my terminal Claude Code is L3, that's why it forgets things"	8
2	没有记忆就没有理解 — Memory as the only durable moat	Memory is what separates L3 from L4	6
3	Agent Memory Architecture — 4-Layer Progressive System	How memory actually works: Context Dir → Progressive Index → FTS5 Recall → Knowledge RAG	10
4	The Six Self-X Properties — what makes agents grow	Self-heal, self-learn, self-improve... the L4→L5 journey	7
5	Coding as Black Box — one requirement in, delivery out	This is how L4 execution works in practice	6
6	Adversarial Review — multi-specialist code review	L4 without quality gates = ships bugs autonomously	8
7	Show your setup — what does context look like on disk?	See a real implementation	5

Optional deep dives:

• The Hidden Defaults That Break Your AI Agent — Claude Code CLI's Undocumented Limits (if you're building on Claude Code SDK — the undocumented maxTurns=100 and task_budget=128K that silently break autonomous pipelines)
• From Zero to Ship in One Session — Building an AI-Ready Engine (if you want to see "Coding as Black Box" in action — one sentence to shipped CLI tool in 4 hours)
• "Dreaming" Is Just Note-Taking — The Real Evolution Happens Elsewhere (if you've heard about Claude Code's "dreaming" and wonder if it's magic — spoiler: it's L0 note-taking with an architectural gap)
• DDD Cultivation — domain knowledge that grows from work (intro) → DDD Full Story — decisions, failures, evidence (deep dive) / 中文版
• Time Symmetry — session boundaries as learning boundaries (if you want smarter session lifecycle)
• Memory Architecture Poll — what's your pattern? (if you're choosing an approach)

After this path, you can:

• Evaluate any AI coding tool by its autonomy level (L1-L5)
• Design a persistent memory layer for your agent (not just in-context)
• Set up adversarial review gates so your agent doesn't ship bugs autonomously
• Articulate why your CLI tool "forgets" — and what architectural layer is missing
• Explain why "dreaming" (background note-taking) is necessary but insufficient for agent learning

---

Path B: Architect

"I'm designing/evaluating agent frameworks"

Step	Article	Why read this	~Min
1	Agent Harness Landscape — different questions, not better answers	The market map — who's solving what	7
2	Agent Harness Autonomy Levels (L1-L5)	Where each product sits, why thickness matters	8
3	Multi-Skill > Multi-Agent — why we chose role-switching	Architecture decision: coordination tax vs context preservation	7
3b	Why We Chose Single-Agent (and multi-agent frameworks prove us right)	CrewAI/LangGraph validation — even "multi-agent" converges to single-process	8
4	Compound Agent Intelligence — design beliefs as compiler guarantees	How to make 1+1+1+1 > 4 architecturally	8
5	Three-Layer Governance — principles, rules, gates	Keep L4+ agents under control without micromanaging	8
6	The Personality Trap — opinionated agents break compliance	The governance failure mode nobody talks about	7
7	Adversarial Review — mechanical quality gates	Trust but verify, structurally	8

Optional deep dives:

• The Hidden Defaults That Break Your AI Agent (if you're embedding Claude Code as SDK — maxTurns, task_budget, autoCompact pitfalls)
• From Zero to Ship in One Session (live case study: DDD + pipeline + adversarial review delivers a CLI tool from one sentence)
• "Dreaming" Is Just Note-Taking (if you're designing agent self-improvement — why L0 memory consolidation plateaus and what L1-L3 looks like)
• Agent Memory Architecture — 4-Layer Progressive System (if you're designing the memory subsystem — FTS5, vectors, distillation)
• AI Agent for Data — from hallucination to precision (if you're connecting agents to business data — semantic contracts, certified queries)
• DDD Cultivation Full Story — decisions, failures, evidence (if you're building domain-aware agents — 3-layer engine, 7 channels, 4 real failures) / 中文版
• No Neo4j — Lightweight Ontology in Practice (if you're choosing knowledge representation)
• Content as Black Box — message in, media out (if you want multi-engine delivery)
• Time Symmetry — session = learning boundary (if you're designing session lifecycle)

After this path, you can:

• Design a harness architecture that separates engine from control plane
• Choose between multi-agent vs multi-skill (and defend the decision with tradeoffs)
• Implement a 3-layer governance model (principles → rules → mechanical gates)
• Design a 4-layer memory system (injection → progressive → FTS5 recall → RAG)
• Spot the compliance failure mode before it ships (personality trap)

---

Path C: Leader

"Should we invest in this? What matters?"

Step	Article	Why read this	~Min
1	Agent Harness Autonomy Levels (self-driving car analogy)	Now I understand what I'm buying/building	8
2	Flat vs Compound — is your AI tool's value curve going up?	The evaluation framework: does it get better with use?	6
3	越用越聪明 — How AI systems compound (and why most don't)	Why 90% of AI tools hit a ceiling	5
4	S×T Tension Matrix — friction = mismatch, not tool quality	Why adoption fails even with good tools	6
5	Four Startup Stages — same stages, different bottlenecks	Where AI changes the game at each phase	5
6	Same Pattern at Every Scale — AI replaces routing, not people	The org design implication	5
7	Stop Selling, Start Showing — "Live Demo of Yourself" > pitch deck	How to actually communicate AI transformation to CXOs	6

Optional:

• AI Agent for Data — from hallucination to precision (if your team is connecting AI to business data — the accuracy/security/scalability tradeoffs)
• Agent 当人来培养 — Cultivation over Configuration (if you manage AI-assisted teams)
• 没有记忆就没有理解 — Memory as moat (if you're evaluating vendor lock-in)

After this path, you can:

• Evaluate any AI tool by asking "is its value curve flat or compound?"
• Diagnose adoption friction using the S×T matrix (is it a supply problem or a tool problem?)
• Make a build-vs-buy decision on agent infrastructure with clear criteria
• Explain to your team why "AI replaces routing, not people" — and what that means for org design
• Pitch AI transformation to CXOs without a deck — using your own experience as proof

---

Full Topic Map (by Theme)

Prereqs legend: → = required (read first or you'll be lost) · ~ = recommended (deepens understanding but not blocking)

Theme 1: Foundations — "What are agents, really?"

#	Title	Lang	Key Idea	Prereqs
38	Design Philosophy — Six Pillars	EN	Prevention, Verification, Knowledge, Correction, Temporal, Ownership	None (START HERE)
39	设计哲学 — 六根支柱	CN	同上中文版	None (START HERE)
33	Agent Harness Autonomy Levels	EN	Harness = control plane, levels = thickness	None
34	Agent Harness 自治五级	CN	同上中文版	None
3	没有记忆就没有理解	Mix	Memory is the only durable moat	None
7	越用越聪明	Mix	Compound > capable — value curves matter	None
6	Agent 当人来培养	Mix	Cultivation > configuration	None
13	Flat vs Compound	EN	Shape of value curves as evaluation tool	→ #7

Theme 2: Architecture — "How do you build one?"

#	Title	Lang	Key Idea	Prereqs
15	Agent Harness Landscape	EN	Market map, different Qs not better As	→ #33
12	Multi-Skill > Multi-Agent	EN	Role-switching beats coordination tax	~ #33
10	Compound Agent Intelligence	EN	Design beliefs as compiler guarantees	→ #7, ~ #8
8	Six Self-X Properties	EN	Self-heal/learn/improve/observe/evolve/compound	→ #3
9	DDD Cultivation	EN	Domain knowledge that grows from work	→ #6
14	Time Symmetry	EN	Session = learning boundary, not idle boundary	→ #3, ~ #8
4	Coding as Black Box	EN	Requirement → verified delivery, no hand-holding	~ #33
5	Content as Black Box	EN	Message → published media, format follows	→ #4
20	No Neo4j — Lightweight Ontology	EN	Darwinian knowledge > graph database	→ #9
19	AI Agent 不需要 Neo4j	CN	同上中文版	→ #9
36	AI Agent for Data — 从幻觉到精准	Mix	Semantic contracts + certified queries for data agents	~ #4, ~ #9
37	Agent Memory Architecture — 4-Layer Progressive	EN	Full memory implementation: injection → progressive → FTS5 → RAG	→ #3, ~ #20
40	DDD Cultivation — Full Story	EN	3-layer engine, 7 channels, 4 failures, real metrics	→ #9, ~ #20
41	DDD Cultivation — 完整故事	CN	同上中文版	→ #9, ~ #19
43	Why We Chose Single-Agent	EN	CrewAI/LangGraph validate single-process	→ #12
45	"Dreaming" Is Just Note-Taking	EN	L0 memory ≠ evolution, L1-L3 is where judgment improves	→ #3, ~ #37
55	From Zero to Ship in One Session	EN	Live case: DDD + pipeline + adversarial delivers CLI from 1 sentence	→ #4, ~ #29
56	The Hidden Defaults That Break Your AI Agent	EN	Claude Code SDK undocumented limits: maxTurns=100, task_budget=128K	~ #4, ~ #33

Theme 3: Governance — "How do you keep it safe?"

#	Title	Lang	Key Idea	Prereqs
26	Three-Layer Governance	EN	Principles → rules → gates (mechanical)	→ #8, ~ #33
27	三层治理模型	CN	同上中文版	→ #8, ~ #33
22	Three-Layer Governance (long)	EN	Full academic treatment with cognitive evolution	→ #26
21	三层治理与认知进化（完整版）	CN	同上中文版	→ #27
31	The Personality Trap	EN	Identity + rules = compliance paradox	→ #26
32	人格陷阱	CN	同上中文版	→ #27
29	Adversarial Review	EN	Multi-specialist review as mechanical gate	~ #4
30	多专家对抗性审查	CN	同上中文版	~ #4

Theme 4: Strategy — "Why does this matter?"

#	Title	Lang	Key Idea	Prereqs
11	S×T Tension Matrix	EN	Friction = Supply × Tool mismatch, not tool quality	None
16	Four Startup Stages	EN	Same stages, AI shifts the bottleneck	None
28	Same Pattern at Every Scale	EN	AI replaces routing, not people	~ #11
46	Stop Selling, Start Showing	EN	"Live Demo of Yourself" > pitch deck	~ #28

Theme 5: Community / Interactive

#	Title	Lang	Type	Prereqs
2	Welcome — What is this gallery?	EN	Announcement	None (START)
17	Memory Architecture Poll	EN	Q&A	#3
18	Show Your Setup	EN	Show & tell	Any
42	100 ⭐ — Thank You (and What's Next)	EN	Announcement	None

---

Dependency Graph (visual)

                    ┌──────────────────────┐
                    │  #33/34 Harness      │  ← BEST ENTRY POINT
                    │  Autonomy Levels     │
                    └────────┬─────────────┘
                             │
              ┌──────────────┼──────────────┐
              ▼              ▼              ▼
    ┌─────────────┐  ┌────────────┐  ┌──────────┐
    │ #3 Memory   │  │ #15 Harness│  │ #13 Value│
    │ as Moat     │  │ Landscape  │  │ Curves   │
    └──────┬──────┘  └─────┬──────┘  └────┬─────┘
           │               │               │
     ┌─────┴─────┐   ┌────┴────┐    ┌─────┴─────┐
     ▼           ▼   ▼         ▼    ▼           ▼
  ┌──────┐  ┌──────┐ ┌──────┐ ┌──────┐  ┌──────────┐
  │#8 6X │  │#9 DDD│ │#12   │ │#10   │  │#7 Compound│
  │Props │  │Cultiv│ │Multi │ │Compnd│  │Intelligence│
  └──┬───┘  └──────┘ │Skill │ │Intel │  └───────────┘
     │               └──────┘ └──────┘
     ▼
  ┌──────────────────────────┐
  │ #26/27 Three-Layer Gov   │
  └──────────┬───────────────┘
             │
      ┌──────┼──────┐
      ▼      ▼      ▼
   ┌──────┐┌──────┐┌──────┐
   │#31/32││#29/30││#21/22│
   │Person││Adver ││Gov   │
   │Trap  ││Review││(long)│
   └──────┘└──────┘└──────┘

---

Quick Reference: All Discussions

#	Title	Theme	Lang
2	Welcome — What is this gallery?	Community	EN
3	没有记忆就没有理解 — Memory as moat	Foundations	Mix
4	Coding as Black Box	Architecture	EN
5	Content as Black Box	Architecture	EN
6	Agent 当人来培养 — Cultivation > Config	Foundations	Mix
7	越用越聪明 — Compound Intelligence	Foundations	Mix
8	The Six Self-X Properties	Architecture	EN
9	DDD Cultivation	Architecture	EN
10	Compound Agent Intelligence — 1+1+1+1 > 4	Architecture	EN
11	S×T Tension Matrix	Strategy	EN
12	Multi-Skill > Multi-Agent	Architecture	EN
13	Flat vs Compound — value curves	Foundations	EN
14	Time Symmetry — learning boundaries	Architecture	EN
15	Agent Harness Landscape	Architecture	EN
16	Four Startup Stages	Strategy	EN
17	Memory Architecture Poll (A/B/C/D/E)	Community	EN
18	Show your setup	Community	EN
19	AI Agent 不需要 Neo4j	Architecture	CN
20	No Neo4j — Lightweight Ontology	Architecture	EN
21	三层治理与认知进化（完整版）	Governance	CN
22	Three-Layer Governance (long version)	Governance	EN
26	Three-Layer Governance (compact)	Governance	EN
27	三层治理模型（简版）	Governance	CN
28	Same Pattern at Every Scale	Strategy	EN
29	Adversarial Review System	Governance	EN
30	多专家对抗性审查系统	Governance	CN
31	The Personality Trap	Governance	EN
32	人格陷阱	Governance	CN
33	Agent Harness Autonomy Levels (L1-L5)	Foundations	EN
34	Agent Harness 自治五级	Foundations	CN
36	AI Agent for Data — 从幻觉到精准	Architecture	Mix
37	Agent Memory Architecture — 4-Layer Progressive	Architecture	EN
38	Design Philosophy — Six Pillars	Foundations	EN
39	设计哲学 — 六根支柱	Foundations	CN
40	DDD Cultivation — Full Story	Architecture	EN
41	DDD Cultivation — 完整故事	Architecture	CN
42	100 ⭐ — Thank You (and What's Next)	Community	EN
43	Why We Chose Single-Agent	Architecture	EN
45	"Dreaming" Is Just Note-Taking	Architecture	EN
46	Stop Selling, Start Showing	Strategy	EN
47	One Builder + AI = One Team. Here's the Proof.	Strategy	EN
49	Your AI Agent Can't "Just Read" a 500K-Line Codebase	Architecture	EN
50	你的 AI Agent 读不懂 50 万行代码	Architecture	CN
51	AI-Ready Repo Standard — Why CLAUDE.md Is Not Enough	Architecture	EN
52	AI-Ready Repo 标准 — 为什么 CLAUDE.md 不够用	Architecture	CN
53	Karpathy's LLM Wiki Is a Manifesto for What We Already Built	Architecture	EN
54	Agent Harness 避坑指南	Architecture	CN
55	From Zero to Ship in One Session	Architecture	EN
56	The Hidden Defaults That Break Your AI Agent	Architecture	EN

---

Quick Stats

• Total discussions: 54 (including Welcome + Reading Matrix)
• Bilingual pairs (EN+CN): 10 pairs (20 discussions)
• English only: 28
• Mixed/bilingual single: 6
• Themes: Foundations (8), Architecture (26), Governance (8), Strategy (6), Community (4)
• Avg reading time per article: ~5-8 min
• Path A total: ~50 min | Path B: ~65 min | Path C: ~41 min