In February 2026, Block CEO Jack Dorsey cut 4,000 employees — roughly 40% of his workforce. Six weeks later, he and Sequoia Capital partner Roelof Botha published "From Hierarchy to Intelligence", a 5,000-word manifesto arguing that AI should replace the entire middle management layer of modern companies.
The essay landed like a grenade. Bloomberg, Fortune, and CoinDesk covered it within hours. The Sequoia co-publication gave it institutional weight that most corporate restructuring memos never get.
But here's the thing nobody in Silicon Valley seems to have noticed: the Tang Dynasty solved the same problem in 618 AD — and their solution lasted 1,300 years because it included a critical component that Dorsey's model is missing.
The Problem Both Are Solving: Human Bandwidth Has a Hard Limit
Before we compare architectures, we need to understand the shared constraint that makes governance necessary in the first place.
The Roman legions discovered it empirically: one person can effectively coordinate 3-8 others. They called it "span of control." Your brain can track what 8 direct reports are doing, what problems they face, what resources they need. It cannot track 50.
So the Romans built hierarchy. Eight soldiers per contubernium, led by a squad leader. Squad leaders reporting to centurions. Centurions to tribunes. The pattern stacked upward because human cognitive bandwidth forced it to.
ℹ️ The key insight: hierarchy didn't emerge because humans love bureaucracy. It emerged because human brains can't process enough information to coordinate at scale without it.
Every organizational innovation since then — from medieval guilds to the East India Company to General Motors — has been a variation on the same theme: how do you coordinate large groups of people when no single brain can hold the full picture?
This is the problem Dorsey is solving with AI. And it's the same problem Emperor Gaozu of Tang solved with three departments and six ministries in 618 AD.
Dorsey's Three-Layer Model: Replace the Router, Keep the Edges
In "From Hierarchy to Intelligence", Dorsey and Botha argue that the core function of middle management is information routing. Managers exist to relay information up and down the chain. If AI can handle the routing, you don't need the routers. Dorsey proposes replacing the entire management chain with three AI layers.
Layer 1: Company World Model — "What's happening inside"
This replaces the manager's function of knowing team status and relaying it upward. Instead of scheduling syncs and writing status reports, an AI system continuously aggregates internal data — code commits, decisions, workflows, performance metrics — to maintain a live picture of company operations.
The prerequisite is that all work must produce digital artifacts. Decisions, discussions, designs, code — everything captured in text. As Dorsey noted, remote-first companies have a natural advantage here because their communication is already text-based.
Layer 2: Customer World Model — "What customers need"
This replaces the traditional PM function of conducting user research, writing PRDs, and deciding the roadmap.
Block has a unique advantage: they own both Cash App (consumer) and Square (merchant), seeing every transaction from both sides. Money is the most honest signal — people lie in surveys but not when spending.
The system observes user behavior and automatically identifies demand signals. Instead of quarterly product cycles, the Customer World Model generates the to-do list in real time.
Layer 3: Intelligence Layer — "What to do now"
This is the decision engine that connects layers 1 and 2. It knows what capabilities the company has (from the Company World Model), what customers need (from the Customer World Model), and automatically composes solutions.
When a combination fails, the failure signal feeds back into the roadmap automatically. Traditional organizations rely on PMs and executives for this judgment call. In Dorsey's model, customer reality generates the priorities directly.
The three layers form an information pipeline:
| Layer | Question | Function |
|---|---|---|
| Company World Model | "What do we have?" | Capability + Status |
| Customer World Model | "What do they want?" | Demand + Signals |
| Intelligence Layer | "What should we do now?" | Decision + Execution |
Together, they replace the entire CEO-to-Manager information processing chain.
ℹ️ From the community: The AI Investor (@The_AI_Investor) summarized Block's restructuring: three roles — ICs who build capabilities, DRIs who own specific problems on 90-day cycles, and player-coaches who combine building with developing people. The Hacker News discussion "Ask HN: When will managers be replaced by AI?" drew 83 comments, with one user noting their company already fired a third of its managers.
The Tang Dynasty Solution: 1,300 Years Ahead of Silicon Valley
While the West was still running simple command hierarchies, the Tang Dynasty (618-907 AD) designed a governance system of remarkable sophistication. The Three Departments and Six Ministries (三省六部) separated planning, review, and execution into independent branches — a separation of powers that Western political philosophy wouldn't formalize until Montesquieu, 1,100 years later.
Three design principles made this system exceptional:
First, separation of planning and execution. The department that drafts policy (中书省) is not the department that executes it (尚书省). This prevents the "I both decide and do" conflict that plagues flat organizations.
Second, mandatory independent review. Every policy must pass through 门下省 before execution. This isn't advisory — it's constitutional. If 门下省 vetoes a policy, it goes back to 中书省 for revision. Even the Emperor couldn't bypass this process without causing a political crisis. This was essentially code review + PR approval, 1,300 years before GitHub.
Third, specialized parallel execution. The six ministries operate independently within their domains. HR doesn't need to coordinate with Engineering for routine operations. This is the same principle as microservices — bounded contexts with clear interfaces.
This architecture outlasted every Western organizational form of its era. It survived because information flowed through checkpoints, not just through chains of command.
The Mapping: Where Dorsey and the Tang Dynasty Align
When you place the two models side by side, the structural parallel is striking:
| Dorsey's AI Layers | Tang Dynasty | Function |
|---|---|---|
| Company World Model | 中书省 (Planning Dept) | Internal state awareness, drafting proposals |
| ??? (Missing) | 门下省 (Review Dept) | Independent review and veto power |
| Customer World Model | 户部 + 早朝官 (Treasury + Intelligence) | External signals, market data |
| Intelligence Layer | 尚书省 + 六部 (Execution + Ministries) | Decision execution, capability composition |
The same governance pattern was independently invented twice — once to govern an empire of 50 million people, once to govern an AI-native company. Both recognized that scale requires separating planning from execution, and that raw information must be processed through structured layers before becoming action.
But there's a critical difference.
The Missing 门下省: Dorsey's Biggest Blind Spot
The Tang system had something Dorsey's model lacks: an independent review layer.
门下省 existed for one purpose — to catch bad plans before they reached execution. If 中书省 drafted a policy that was contradictory, poorly thought out, or harmful, 门下省 could kill it. No exceptions.
In Dorsey's three-layer model, information flows from Layer 1 to Layer 2 to Layer 3 with no independent checkpoint. Who audits the Company World Model for accuracy? What happens if the Customer World Model misreads a signal? What if the Intelligence Layer composes a capability combination that looks optimal by the numbers but creates a terrible user experience?
⚠️ This is exactly the problem that killed previous flat organization experiments. Zappos' Holacracy, Valve's no-manager structure, Spotify's Squad model — they all collapsed because removing hierarchy also removed the implicit quality checks that managers provided. The hidden power structures just went underground.
The Tang Dynasty solved this explicitly. 门下省 wasn't optional — it was structural. Conflicts between departments were resolved at the review stage, before execution, not after. This is why the system lasted 1,300 years while Silicon Valley organizational experiments typically last 3-5 years before quiet reversion to hierarchy.
The implication for AI-native organizations is clear: if you're building an AI system that plans and executes, you need a separate AI system that reviews and can veto. The reviewer cannot be part of the planning or execution chain. It must be structurally independent.
The Open Source Response: Edict and Multi-Agent Governance
The developer community has already noticed this gap. An open-source project called Edict (6.9K+ stars on GitHub) explicitly implements the Three Departments and Six Ministries architecture as a multi-agent orchestration system.
Edict runs 9 specialized AI agents, with the critical innovation being a dedicated review agent modeled after 门下省. This agent can reject outputs from planning agents and force them to redo work — a structural veto power that mainstream multi-agent frameworks like AutoGen and CrewAI lack.
The contrast is instructive:
| Framework | Approach | Review Mechanism |
|---|---|---|
| AutoGen / CrewAI | Agents "chat freely" | None — agents self-organize |
| Dorsey's Block Model | Three AI layers | None — information flows one-way |
| Edict (三省六部) | Structured pipeline | Dedicated review agent with veto |
Most current multi-agent frameworks let agents self-organize through conversation. This works for simple tasks but breaks down at scale — agents hallucinate, enter loops, drift off-topic, and there's no structural mechanism to catch these failures. Edict's approach treats review not as a plugin but as architecture. As we explored in our look at the rise of AI agents, the agent landscape is evolving fast — and governance is becoming the key differentiator.
The broader conversation is exploding across platforms. On Substack, Ken Huang asks "What is an Agentic AI Native Organization?" — arguing that once AI becomes agentic, the org chart becomes an execution graph. GAI Insights cuts through the buzzwords: the real metric isn't whether you're "AI-native" or "AI-first" — it's revenue per employee.
What This Means for Practitioners
Whether you're restructuring a company or building a multi-agent system, the convergence of these ideas points to three actionable principles:
1. Make All Work Machine-Readable
Dorsey's Company World Model requires digital artifacts for everything. This is the foundation — without it, nothing else works.
Practically: put an AI scribe in every meeting. Move key decisions from hallway conversations to documented channels. Make Slack, docs, and code your single source of truth. Remote-first companies already have this. Office-first companies need to build it deliberately.
2. Separate Your "Information Routing" from Your "Real Judgment"
List everything your managers do weekly. Categorize each task:
| Function | % of Time | AI Replaceable? |
|---|---|---|
| Status syncs, info relay | ~40% | Yes — fully automatable |
| Priority setting, resource allocation | ~20% | Mostly — AI can optimize |
| Quality review, risk assessment | ~10% | Partially — AI assists, humans decide |
| Conflict resolution, politics | ~10% | No — requires power and trust |
| People development, coaching | ~10% | No — requires human connection |
| Culture, values, morale | ~10% | No — requires authenticity |
AI can replace roughly 60% of what middle managers do today. The remaining 40% — judgment, people, culture — stays human. This doesn't mean you need fewer managers. It means management becomes a behavior (something senior ICs do part-time) rather than a full-time role. For a deeper look at how these AI tools actually work in practice, see our guide to what AI agents are and how they work.
3. Add a "门下省" to Your AI Systems
Whether you're designing an AI-native organization or building a multi-agent pipeline, ask: who reviews the AI's output before it reaches the customer?
This reviewer must be:
- Independent — not part of the planning or execution chain
- Empowered — able to reject and force revision, not just flag concerns
- Structural — built into the architecture, not bolted on as an afterthought
The Tang Dynasty figured this out 1,300 years ago. The most sophisticated bureaucracy in human history survived because review wasn't optional — it was constitutional.
The Meta-Insight: Governance Is Eternal
Put Dorsey's AI-native organization and the Tang Dynasty's Three Departments together, and you get a closed narrative loop:
- Human hierarchy is too slow → Dorsey proposes replacing it with AI information routing
- But AI agents also need coordination → multi-agent frameworks emerge
- Multi-agent frameworks without governance fail → the Tang Dynasty's review architecture resurfaces
- Therefore: hierarchy doesn't disappear, only its medium changes
Whether the coordinating entities are carbon-based (humans) or silicon-based (AI agents), the fundamental governance challenges remain the same: how do you plan, review, execute, validate, and balance at scale?
Dorsey's contribution is asking the right question: if AI can handle information routing, do we still need the current organizational structure? The Tang Dynasty's contribution is the answer the question implies: you still need structure — just make sure it includes independent review.
The answer for 2026 probably won't look exactly like Dorsey's vision, and it won't be a perfect replica of the Tang Dynasty system. But the core insight from their collision is worth remembering:
💡 Governance problems are eternal. The medium changes — from humans to AI to whatever comes next. But the pattern holds: plan, review, execute, validate, balance. This has been true for 1,300 years. It will be true for the next 1,300.
Sources: Block — From Hierarchy to Intelligence | Sequoia Capital | Fortune | Deloitte Tech Trends 2026 | Edict on GitHub | Growth Unhinged | HBS Online
Originally published at ComputeLeap
Top comments (0)