Chapter 10: The Collapse of the Execution Layer: Evolving the Talent Structure

The default value of one or two thousand years is questioned by a letter #

In February 2026, payment company Block laid off about 4,000 people, nearly 40% of its total employees. The reason founder Dorsey gave on social media was not market headwinds, but an organizational statement: "The intelligent tools we are creating and using, coupled with smaller and flatter teams, are bringing about a new way of working that fundamentally changes 'what it means to build and run a company.'" A month later, he and Sequoia partner Botta jointly published an article with the title of this chapter: "From Hierarchy to Intelligence."

The sharpest thing in the article is a piece of historical archeology. In a Roman legion, eight soldiers in a tent needed a captain, eighty men needed a centurion, and five thousand men needed a legionnaire. For two thousand years, from Roman military camps to today's multinational corporations, the default value of organizations has never changed: when there are more people, layers must be added, because people need to carry information between layers, precalculate decisions, and maintain alignment. Dorsey and Botha's questions hit the very center of this default value: The question has never been whether these layers are needed, but whether humans are the only option to assume these functions. It was true for two thousand years, but no longer. **

Chapter 1 talked about the execution price approaching zero, and Chapter 9 talked about context being an organizational asset. Twist the two threads together and that's this chapter: Hierarchy is the old infrastructure of transporting context with human flesh. When AI takes over execution and the "world model" takes over context handling, not only the underlying execution posts will collapse, but also the entire middle layer that handles information. What the organization's talent structure faces is not optimization but redrawing.

Let’s talk about the middle layer separately, because it is the most easily missed group of people in the collapse zone. If you look at a typical middle-level manager’s week, he spends most of his time on three things: upward summarization (processing subordinates’ progress into the format required by superiors), downward translation (breaking down superiors’ intentions into subordinates’ tasks), and horizontal alignment (cross-department meetings to synchronize information). The common essence of the three things is context transfer, and transfer is exactly the job of the "world model" in the previous paragraph (this word comes from Block, expanded in the third section). Note that this judgment is completely different from "management is of no value": middle-level positions are usually mixed with two elements, one is handling, and the other is judgment: coaching people, handling exceptions, and making choices in ambiguous areas. Collapse absorbs the former and leaves behind the latter. So there is not one but three ways out for the middle level. For those with a high judgment component, move to player-coach: continue to coach people, but use judgment guidance instead of information gap guidance. If you have deep expertise, go back to individual contributors: in Block’s structure this is not a downgrade, it’s a change to a more expensive approach. The person whose two components are thin and whose position is entirely composed of transportation is the one who truly stands in the middle of the collapse zone. An honest organization should expose all three paths instead of leaving all middle managers to guess and defend.

2. Who is standing on the collapse zone: S-class classification #

First, let’s clarify “which positions are in the collapse zone”. Our company uses a set of S-levels to grade positions. This chapter borrows them as a ruler:

S1, follow instructions: given goals, steps, and deadlines, follow them. Data entry, basic customer service, templated copywriting and reports.
S2, proficient execution according to standards: given standards, completed with proficient skills, no need to re-judge. Primary code implementation, regular design delivery, and process-based operation.
S3 and above, judgment is required: making choices within the standard, handling exceptions not covered by the standard, and participating in the iteration of the standard itself.

The collapsed judgment line is drawn between S2 and S3. What S1/S2 have in common is that the judgment has been done by someone else. Precisely because judgment is put forward, these positions can be written as standards (Chapter 6), and jobs that can be written as standards can be fed to AI or outsourced. This is the mechanism derivation of the assertion in this chapter. There is no need to predict which model will be released in which year: **The closer a position is to "standardizable", the closer it is to the collapse zone. **

This leads to the negation of a popular move. When AI comes, the first reaction of most companies is to provide "AI empowerment training" to employees and teach them how to use the tools. There is nothing wrong with the training, but it is in the wrong direction: the problem is never whether employees can use tools, but whether the job itself is viable or not. No matter how proficient an employee in an S1 position is in using AI, he is only accelerating the standardization process of his position. **What the organization really needs to do is to redraw the talent structure, not to hold training courses for the old structure. **## 3. An experimental machine in live broadcast: three roles of Block

Block published the redrawn structure directly. The entire company falls into three roles:

Individual Contributor: Deep experts, key changes in information sources, they get context from the company's "world model" rather than from managers. A layer of intelligence is built on all work traces (messages, emails, code submissions, documents, meeting recordings) to continuously maintain a real-time model of the entire company, and anyone can "talk to the company" directly. Directly Responsible Person: There is only one person responsible for everything, and he is responsible for the results. Player-coach: Train the people around you while doing business; this role replaces the traditional manager whose job is to transfer information: he also writes code and designs, but no longer acts as a human relay for context. The structural goal given by Dorsey: The level between the CEO and any employee will be reduced to two or three levels this year. The ideal state is zero level: six thousand people report directly, "because most of the work has flowed through the intelligent layer, this matter has become manageable."

Citing this case, I will first describe the boundaries: Block's transformation is in the early stages. Dorsey and Botha admitted in the original article that "it will be very difficult, and some links may break first." Therefore, I only quoted its mechanism design and did not draw any conclusions about success or failure. Its value lies in providing a complete new structural drawing that is actually running: a small and all-staff thinking core, plus the outer circle of AI and outsourced execution belts**. This is also the framework diagram for this chapter. As for whether the drawings can be built into a building, I will update it honestly in subsequent versions of this book.

Block is a drawing of six thousand people, and the verification cycle is measured in years. Pair it with a small sample that has already run a full circle: our own R&D unit. The scale is small enough to be transparent, just enough to spread all the accounts. In March 2026, a peer survey triggered the decision to restructure (alumni feedback was that "interns have used it well", ask who has done it before starting); the R&D executive at the time was two skilled engineers. Four months later, the new structure was running stably: AI was responsible for the execution body, and nine interns formed a mobile execution belt. It turned out that one of the two was transferred to a judgment position: from writing code to directing interns, setting standards and controlling output, which is exactly what Block said as player-coach. Another person left his job. The account books are as follows (company operating standards, unaudited): monthly labor costs have increased from about 35,000 yuan to about 80,000 yuan, about 2.3 times, including the cost of AI tools; production capacity measured by iteration speed and functional output is about five times that of the past; R&D response speed is about three times; the BUG rate in the production environment has dropped by about 80%. Please note the shape of this ledger: Collapse is not a money-saving exercise, it is the same budget spent in a different way. Costs are rising, and the cost per unit of production capacity has dropped to less than half of its original value.

Both sides of this sample should also be accounted for. Quality declined in the first month: nine interns were brought in, the execution team started running first, and the standards did not keep up. The corrective action is not to withdraw people, but to go back and make up for the lessons in Chapter 6: establishing a quality and standards system. The decrease in BUG rate occurs after the standard system is established**. This sequence is the mechanism of the conclusion in this chapter: interns plus explicit standards outperform skilled engineers plus tacit experience; conversely, without that set of standards, nine interns are nine times more chaotic. The training system is still incomplete. This is the part of this sample that has not yet been settled.

4. Those left behind: Everyone is a future partner #

Let me first clarify a question left in the previous section: the interns and AI on the execution team are mobile capabilities, not positions. They scale with the scale of the task and do not occupy the organization's judgment structure. Every real position left in the new structure is above S3, which changes two things: who to recruit and how to invest.

Our internal version of recruitment criteria is four: unity of values, strong thinking ability, self-driving force, and ultimate execution. Note that execution is still on the list. What has collapsed are pure execution positions, not execution capabilities; S3+ people must be able to both judge and act. This is what Block's player-coach "also writes code" means. All four together point to an acceptance question, which is the only question I really ask myself when recruiting: Is this person worthy of being my partner in five years? If the answer is that they are not qualified, then if you ask a little deeper, you will usually find that the position itself should not be filled by people.These four items are not hung on the wall on the first day you open a company. Before they existed, our recruitment was different: without standards, we could only relax requirements, recruit in large quantities, and then eliminate people after they came in. We used funnels to hedge against the absence of judgment, and postponed screening from the interview stage to the probation period. The cost was paid by the company and the people who were eliminated. There is a specific figure for this tuition fee: nearly twenty people were recruited before and after the business position, but not a single suitable one was left. No matter how wide the funnel is, absent standards will not grow on their own. This is a standard form of unstandardized execution in recruitment (the iron law in Chapter 6, we violated it ourselves first), and it wastes far more than recruitment fees: every person eliminated has consumed a pit, management bandwidth and several months of two-way time. In April 2026, after the four standards were introduced, the first thing we did was not to use them to screen new employees, but to go back and re-evaluate all current positions and calibrate the responsibilities and performance standards of each position according to the new standards. This sequence is worth writing down separately: The first batch of users of the standard will always be existing users, not incremental: Only screening new people and not recruiting old ones is tantamount to announcing that the standard only applies to people without voting rights.

How to invest in people, the ledger from Chapter 5 is directly transferred: each person is a human investment that must prove a positive cycle within the stage window, the personal ROI is greater than 1, and the acceptance is three months in the early stage, six months in the middle stage, and nine to twelve months in the later stage. In the executive era, "mediocre but obedient" can be tolerated because execution itself has output; in the S3+ structure of all employees, a person who does not produce incremental judgments occupies the most expensive position in the entire organization.

5. Ladder of decision-making power: from instruction to delegation #

Once people are recruited, how will judgment grow? Rely on a five-level ladder of decision-making power. It does not describe job promotion, but rather that managers will transfer their decision-making power to one level after the members’ judgment matures:

Instructions: Managers give goals, standards, steps and deadlines. Suitable for new tasks, new members, and high-risk scenarios.
Guidance: Managers clarify goals and standards and explain key methods; members complete and review under guidance.
Negotiation: Managers only provide context and boundaries; members propose solutions, and both parties make joint choices and commit to each other.
Authorization: Managers clarify results, authority, resources and risk boundaries; members make decisions independently within the boundaries and are responsible for the results.
Decentralization: Managers only maintain the mission, values and a few red lines; members are also responsible for problem definition, resource allocation and standard iteration.

The difference between delegation and delegation deserves a separate line: delegation means making independent decisions within established boundaries, while delegation means leaving even the optimization of boundaries to the other party. The prerequisite for each level of migration is that the infrastructure in the previous chapters is in place: the standard has been established (Chapter 6) and the context has been understood (Chapter 9). Only then can the maturity of the judgment be tested. Delegating power without standards is not trust, but loss of control; conversely, when standards and context are all in place, everyone is still at the "instruction" level. This is the old gene that this chapter will replace: on the side of talents who perform worship, they only reward obedience and do what they are told, while judgment shrinks while waiting for instructions.

Here we must first get rid of a mild illusion: from S2 to S3 is a slope, and you can walk up it by working a few more years and learning a little more. no. There is a gap between S2 and S3 (borrowing Moore's word, because the structure is completely isomorphic): the two sides of the gap want different things, and the kinetic energy on the left is accumulated to the edge and is cleared. Specifically, a quadruple fracture. The axis of assessment has been changed: The excellence of S2 is the fast and accurate implementation of the standards, while the excellence of S3 is making the right choices in areas that are not covered by the standards. Practicing and executing for 10,000 hours will not automatically produce a gram of judgment. Habits are exactly the opposite: S2’s virtues (not questioning standards, zero deviation delivery) are exactly S3’s obstacles (questioning standards, exposure to risk). To cross the chasm, you must first uninstall some of the habits that have made you successful so far. The feedback structure has changed: The feedback on execution is fast and clear, and the feedback on judgment is slow and vague, mixed with luck. People without attribution tools fall into this feedback vacuum and fall back into the safe zone after making the first misjudgment. Responsibility jump: An error in S2 is an execution accident, an error in S3 is because your judgment was wrong, and you need to sign. Many people are not incompetent, but they are unwilling to sign for the results and stay on the left bank of the chasm their entire lives.

Seeing these four fractures clearly reveals the true identity of the decision-making power ladder: it is the organization's bridge to the S2-S3 gap. The fifth level is not five titles, but five sections of bridge: for each level, one level of judgment responsibility is added, and one level of safety net is removed. At the directive level, all judgments are made by the manager. At the guidance level, members are allowed to see how the judgment is made. At the negotiation level, he is asked to make the judgment but both parties share the responsibility. At the authorization level, he is asked to sign but the boundaries are guaranteed. At the delegation level, even the boundaries are handed over. The chasm cannot be crossed in one step, but it can be crossed in sections. Organizations cannot wait for employees to grow wings on their own, they must build bridges; employees should not expect to be able to slide across them after a long time, and they must actually walk on each section of the bridge.The acceptance of the transfer must also be clearly stated, otherwise the ladder will degenerate into another set of annual merit sequence with automatic promotion based on seniority. The signal to move up to a higher level is to have both: consecutive results at the current level** (evidence to judge quality), plus active exposure to risks or bad news at least once (evidence to judge honesty). The second signal is easy to ignore, but it tests the most difficult level of the fourfold fracture: whether you are willing to sign for the judgment. The reverse is also true: downgrade is not punishment, but calibration: for a person who has crossed the boundary twice in a row at the authorization level, it is much cheaper for him and the organization to return to the consultation level and walk a section of the bridge again than to fall down for the third time.

If you are standing on the left bank of the chasm at the moment, and the bridge has not yet reached your feet, there is a way to cross the bridge that does not require waiting for an organization: Write standards for the work you are performing. The act of writing standards forces you to make judgments: what counts as good, where the boundaries are, and what to do with exceptions. They are all S3 muscles. Moreover, it is extremely low-risk. No one will be hurt if it is written incorrectly, and it will be seen immediately if it is written correctly. The second step is to proactively claim exceptions that are not covered by the standards. Exceptions are a natural exercise in judgment. The place where others avoid is your bridge. The third step is to use the three questions in Chapter 11 to build an attribution loop for yourself, turning the right or wrong of each judgment into accumulated experience. None of the three steps requires a title, which is exactly the point: the fastest S3 ticket is to write standards for the work you are performing: judgment is never awarded, it is ratified after being practiced. **

Boundaries of responsibility on this side of the organization need to be drawn with the same precision. The bridge is built by the organization, but the bridge does not carry people's back. When it comes to crossing a bridge, the mechanism can only be screened but not cultivated: in the four-fold fracture, there is no training that can be injected from the outside in terms of unloading old habits and being willing to sign for the results. Therefore, all the actions of the organization are a three-piece set: establishing standards (write down and make public the "good" observable characteristics of each level, and those being observed must know what to observe in advance), opening channels (adding the bridge deck of the stairs to write standards, claiming exceptions, and three questions, and these bridge crossing methods are open to everyone), establishing observations (for promotion and demotion, only behavioral records are recognized, and self-reports and impressions are not recognized). Evaluation must be observation, because judgment cannot be measured verbally: interviews measure self-reports, job descriptions are performances, declarations are used to manage other people's judgments, and behavioral records reveal the true level. This is completely the same as evaluating AI. No one can open the model to confirm whether it "understands" it. They can only design tasks, observe output, and score against standards. We actually all have observation devices in our hands (ROI cards, standard output counts, exception handling records, Feynman acceptance), but what we lack is usually the awareness to use them as observation devices. There is also an easy pitfall buried here: Position names do not carry hierarchical information. With the same "HR" title, those in ordinary companies execute the recruitment process in S2, while those in top companies design talent standards and observation systems in S4/S5. The functions are the same, but there are two or three levels of judgment; using functional names to manage the talent structure is equivalent to measuring people with a ruler without scale.

The role of managers on this ladder has also been redefined: The end point of management is not to make subordinates more obedient, but to allow more people to have reliable judgment rights: transform yourself from a decision-making bottleneck to a distributor of decision-making power. This is exactly Block's player-coach, which is the same thing in our language.

6. Loser sample: companies with locked-up decision-making power #

If decision-making power is not transferred, people will die. Ofo has demonstrated this with a valuation of US$4 billion.

The mechanism is deeper than the popular "Dai Wei is too stubborn". The true terms after verification are: ofo's board of directors stipulates that financing and other major matters require the signature of all shareholders. From October to November 2017, during the window period for the merger with Mobike, Dai Wei used this mechanism to veto the merger. His original words were "Even if there is a merger, we will merge with Mobike." He even rejected the compromise proposal of "becoming the sole CEO of the merged company in six months." Then this clause began to backfire: the same "all signature" mechanism was used in reverse by Didi and Alibaba: for more than a year after that, no new financing could be finalized, and the only thing that could pass was a mortgage loan, because that only required more than half of Dai Wei's voting rights. The company sank in a two-way paralysis: the founder locked up the shareholders, and the shareholders locked up the company. The end result is settled by the user: At the end of 2018, there were more than 10 million users queuing up online to refund their deposits, and the queuing numbers were in the tens of millions. That string of numbers is the most intuitive picture of the five words "judgment cannot flow".

Analyzed in the language of this chapter, this is the complete opposite of the decision-making power ladder: Provisions designed for defense lock the judgment migration of the entire organization. Judgment power can neither go down (all major judgments within the organization converge on the founder), nor can it go out (the terms make it impossible for any external force to correct the deviation). The control group is Didi and Kuaidi: before the merger, the veto rights of investors were collectively removed, and similar clauses were actively removed before a successful merger. The two cases put together establish an iron rule for the design of decision-making rights: **Any clause that prevents the flow of judgment, no matter who is originally designed, will ultimately trap the organization itself. **

By the way, let’s align our own progress: the complete account book of the R&D unit has been spread out in the third section; the other execution posts are following the same transformation route, and the effect is similar to that of the R&D unit, but the sample size is not as clean as the R&D unit, so I chose not to spread it out; the full case in Chapter 13 will summarize the assembly drawings of all units.

7. Boundaries of JudgmentThree. #

First, collapse is a direction judgment, not a layoff order. Collapse rates vary by an order of magnitude across industries (the boundaries from Chapter 1 apply here as well), and physical delivery-heavy business S1/S2 jobs will be around for a long time. The message to employees in this chapter is not "you are eliminated", but a clear route across the river: S2 to S3 is a chasm, not a ramp (Section 5). The pier on the individual side is the four actions of adaptive insight in Chapter 3 plus the two learning tools in Chapter 7. The bridge on the organizational side is the ladder of decision-making power. For individuals in the collapse zone, the way out is written in the previous chapter; what the organization owes its employees is to build the bridge.

Second, "Everyone is a partner" is the structural direction, not tomorrow's equity plan. It talks about the convergence of the nature of positions: when every position left bears judgment, the relationship between the organization and its members will inevitably evolve from hired labor to partnership judgment. How to keep up with equity, distribution, and governance is each company's own engineering problem, and this chapter does not provide a prescription.

Third, tell the truth about the quality of the sample. Block is in the early stages of transformation and may partially collapse, so I will only cite the mechanism design; ofo's "all signature" mechanism has been cross-corroborated by multiple reports, but there are contradictions in the statements of the parties on the veto list, so I will only cite the mechanism and not the list; the background of the recruitment standards (Section 4) and the account books reconstructed by the R&D unit (Section 3) are both company operating standards and have not been audited, so they are discounted based on the self-report. There are still two samples missing: a case of a traditional enterprise that successfully transformed its talent structure, and a loser that "used AI to replace hierarchical coordination and failed." The second one is particularly important. Block is now an isolated certificate, and the isolated certificate can only be used as a hypothesis.

What to Do Monday Morning (No. 1 perspective) #

Three steps, within two hours:

Mark S Level: Mark each position on the organizational chart with S level. Don't ask "Can this person be replaced by AI?" Ask "Is the judgment of this position already done by others?" The answer is "yes", it is S1/S2. Count the S3+, that's what's left after your organization collapses. If the ratio is less than 30%, redrawing of the talent structure will be put on the agenda.
Mark management level: Mark the current level of decision-making power (instruction/guidance/consultation/authorization/delegation) for each core member. Then ask a question: If he can't be promoted, is it lack of understanding of standards, sustained results, or the ability to honestly expose risks? The answer determines which lesson you should teach him, and also tests whether you have bottlenecked your decision-making power.
Check the clauses: Dig out the company's articles of association and investment agreement, find all the clauses that "require unanimous consent" or "unilateral veto", and ask one by one: Will this clause prevent the flow of judgment in the worst-case scenario? Ofo's lesson is that such clauses are invisible at ordinary times and unsolvable when they occur.

Side note (individual and team perspective): Give yourself an S grade, the same criteria as above: How much of the judgment of the work at hand has been done by others? The higher the proportion, the more time should be allocated to practicing the four movements. The collapse zone is not scary. It is scary to stand in the collapse zone without knowing it.

Connections to Adjacent Chapters- Continuing from Chapter 9: What kind of people are needed to run the consensus engine. The old gene replaced in this chapter is "Execution Cult" (talent side) - only rewards obedience and doing things. #

Handed over to Chapter 11: How should those who stay spend their day? ——Time matching.

Chapter Acceptance Self-Check (compare with the five acceptance standards of the chapter) #

The assertion can be restated in one sentence ✓, and it is the inference of the core assertion (execution returns to zero → the positions whose profession is execution and the hierarchy whose profession is transporting information collapse at the same time).
Whiteboard framework diagram ✓ (pyramid → small core + AI/outsourcing execution belt morphological change diagram + five-level decision-making power ladder).
External comparison and data ✓: Positive Block (newly verified on 2026-07-26: layoff data, original text of the declaration, three-role mechanism; only design is cited, not success or failure); loser ofo ("all signature" mechanism has been verified, Didi Kuaidi's comparison); R&D unit reconstruction has been written into Section 3, The recruitment standards and background have been written into Section 4 (2026-07-26, company-reported basis, unaudited, quality has been stated in the judgment boundary); the two missing samples have been truthfully marked.
21 golden sentence candidates ✓ (v1.4 adds two sentences about the double components of the middle level and S3 admission ticket; v1.5 adds three new sentences about bridges do not back people, observation, and job titles).
"What to Do Monday Morning" Three steps from the No. 1 perspective + personal notes ✓.