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

1. A two-thousand-year default, questioned by a letter #

In February 2026, the payments company Block laid off about four thousand people—nearly forty percent of its workforce. Founder Jack Dorsey gave the reason on social media as an organizational statement: the intelligent tools they were building and using, together with smaller and flatter teams, were producing a new way of working that fundamentally changed “what it means to build and run a company.”

A month later he and Sequoia partner Roelof Botha published a joint essay whose title is this chapter’s theme: From Hierarchy to Intelligence.

The sharpest move in the essay is a piece of historical archaeology. In a Roman legion, eight soldiers sharing a tent needed a squad leader, eighty men needed a centurion, five thousand needed a legion commander. For two thousand years, from Roman camps to today’s multinationals, the organizational default never changed: more people meant more layers, because someone had to carry information between layers, precompute decisions, and keep alignment. Dorsey and Botha’s question lands at the center of that default: the question was never whether those layers are needed, but whether humans are the only option for those functions. For two thousand years, yes. Not anymore.

The price of execution is approaching zero, and context is an organizational asset. Twist the two threads together and hierarchy is the old infrastructure for hauling context with human bodies. When AI takes over execution and a “world model” takes over context hauling, what collapses is not only the bottom execution belt but the entire middle layer whose job is to move information. The talent structure needs to be redrawn.

The middle layer is the group most easily undercounted in the collapse zone, and it deserves its own treatment. Break open a typical middle manager’s week and most of it goes to three things: rolling up, translating down, and aligning sideways. Rolling up means turning subordinates’ progress into the format superiors want. Translating down means breaking superiors’ intent into tasks. Aligning sideways means cross-team meetings to sync information. All three are, at root, context hauling—exactly the job of the “world model” in the paragraph above (Block’s term).

Middle roles usually mix two ingredients: hauling is one; judgment is the other—coaching people, handling exceptions, making trade-offs in the fog. Collapse sucks away the former and leaves the latter. Management still has value. What collapses is the hauling.

So the exits are three, not one. Where judgment is thick, move toward player-coach: keep leading people, but lead with judgment rather than with an information gap. Where craft is deep, return to individual contributor—in Block’s structure a different way of being expensive, not a demotion. Where both ingredients are thin and the role is pure hauling, that person stands in the middle of the collapse zone. An honest organization puts all three paths on the table instead of leaving every middle manager to guess and defend.

2. Who stands in the collapse zone: the S ladder #

Which roles sit in the collapse zone has to be named precisely. Internally we grade roles on an S ladder; this chapter borrows it as a ruler:

• S1, execute on instruction: goals, steps, and deadlines given—follow them. Data entry, basic support, templated copy and reports.
• S2, execute proficiently to a standard: a standard is given; skilled hands finish it without reopening the judgment. Junior implementation, routine design delivery, process operations.
• S3 and above, judgment required: trade-offs inside the standard, exceptions the standard does not cover, and participation in iterating the standard itself.

The collapse line is drawn between S2 and S3. What S1 and S2 share is that the judgment has already been made by someone else. Precisely because judgment is front-loaded, those roles can be written as standards—and work that can be written as a standard can be fed to AI or outsourced. That is the mechanism behind this chapter’s claim, and it does not require a forecast of which model ships which year: the closer a role is to “standardizable,” the closer it sits to the collapse zone.

When AI arrives, most companies’ first reflex is “AI enablement training”—teach people the tools. Training is fine; the direction is wrong. The question was never whether employees can use tools, but whether the role itself should survive. An S1 employee who becomes fluent with AI is only accelerating the standardization of their own role. The organization must redraw the talent structure. Training classes for the old one will not do that job.

3. A live experimental machine: Block’s three roles #

Block published the redrawn structure outright. The whole company maps to three roles.

Individual contributors are deep experts. The key change is the source of information—they take context from the company’s “world model,” not from a manager. An intelligence layer sits on every work trace (messages, email, commits, docs, meeting recordings) and maintains a live model of the company; anyone can “talk to the company” directly.

The directly responsible person is the single owner of each matter, accountable for the outcome.

The player-coach does the work while developing the people nearby. This role replaces the traditional manager whose profession was hauling information: they still write code and still design, but they no longer act as a human relay for context.

Dorsey’s structural target: layers between CEO and any employee down to two or three this year; the ideal is zero. Six thousand people reporting directly, “because most of the work already flows through the intelligence layer, and that becomes manageable.”

Block’s transformation is early. Dorsey and Botha say in the essay that it will be hard and that some links may break first—so this chapter cites only the mechanism design, never a verdict of success or failure. Its value is a complete new structural drawing that is actually running: a small core in which everyone thinks, plus an outer belt of AI and outsourced execution. Whether the drawing becomes a building is for time to settle.

Figure: after collapse, a judgment core remains; rights migrate as judgment matures.

Block is a six-thousand-person drawing; the verification cycle is measured in years. Pair it with a small sample that has already run a full loop: our own R&D unit. Small enough to be transparent—exactly the right size to open every ledger.

In March 2026 a peer survey triggered the rebuild. Alumni said “interns worked well”; we asked who had done it before we moved. The execution layer then was two skilled engineers. Four months later the new structure was stable: AI carried the execution body; nine interns formed a mobile execution belt. Of the original two, one moved onto a judgment seat—from writing code to directing interns, setting standards, and gating output, exactly Block’s player-coach. The other left.

The ledger (company operating figures, unaudited): monthly people cost from about RMB 35,000 to about RMB 80,000—roughly 2.3×, including AI tools; capacity by iteration speed and feature output about five times prior; R&D response about three times; production bug rate down about 80%. Absolute cost rises; cost per unit of capacity falls to less than half. Collapse spends the same budget a different way.

Both sides of the sample must be told. Quality fell in month one: nine interns came in, the execution belt started first, standards lagged. The correction was not to pull people out, but to go back and build a quality-and-standards system. The drop in bug rate happened after that system stood up. That sequence is the claim’s mechanism in the field: interns plus explicit standards beat skilled engineers plus tacit experience. Without those standards, nine interns are nine times the chaos. The training system is still incomplete. That part of the sample has not yet settled.

4. Those who remain: everyone a future partner #

Interns and AI on the execution belt are mobile capacity, not roles. They scale with the task and do not occupy the organization’s judgment structure. Every real role that remains in the new structure sits at S3 or above. That changes two things: whom you hire, and how you invest in people.

Our internal hiring standard is four lines: shared values, strong thinking, self-drive, and extreme execution. Execution stays on the list. The pure execution role collapsed; execution capability did not. S3+ people must both judge and do; Block’s player-coach “still writes code” is that idea. The four lines converge on one acceptance question, and the only question I truly ask myself when hiring: in five years, is this person fit to be my partner? If the answer is no, dig one layer deeper and you usually find the role itself should not be filled by a human.

Those four lines were not on the wall on day one. Before them, hiring looked different: no standard, so we loosened requirements, hired in bulk, and culled after entry. We used a funnel to hedge the absence of judgment, postponing screening from interview to probation, with cost paid by the company and by those cut.

The tuition has a concrete number: nearly twenty people hired for a business role, and not one suitable person remained. Widen the funnel as much as you like; a missing standard will not grow itself. That is the standard shape of execution-without-standards in hiring—we violated it ourselves first—and the waste is far more than recruiting fees. Every person cut consumed a seat, management bandwidth, and months of two-way time.

In April 2026, once the four standards were derived, the first act was not to screen newcomers with them, but to re-grade every sitting role and recalibrate duties and performance by the new standard. The order cannot reverse: a standard’s first users are always the stock, never the increment. Screen only newcomers and leave old roles untouched, and you announce that the standard applies only to people without a vote.

How to invest in people: bring the ledger straight over. Each person is a human investment that must prove a positive loop inside a stage window—personal ROI greater than one; early stage three months, mid six, later nine to twelve. In the execution-role era you could tolerate “mediocre but obedient,” because execution itself produced output. In an all-S3+ structure, a person who adds no judgment increment occupies the organization’s most expensive seat.

5. The decision-rights ladder: from instruction to delegation #

People are hired—how does judgment grow? On a five-rung decision-rights ladder. The ladder tracks judgment maturity, not title promotion: as a member’s judgment matures one level, the manager migrates decision rights one level.

1. Instruction: the manager gives goals, standards, steps, and deadline. For new tasks, new members, high-risk scenes.
2. Guidance: the manager clarifies goals and standards and explains key methods; the member finishes and reviews under guidance.
3. Consultation: the manager supplies only context and boundaries; the member proposes options; both choose together and commit to each other.
4. Authorization: the manager clarifies outcomes, authority, resources, and risk bounds; the member decides inside the bounds and owns the result.
5. Delegation: the manager keeps only mission, values, and a few red lines; the member also owns problem definition, resource allocation, and standard iteration.

Draw the line between authorization and delegation: authorization is independent decision inside a set boundary; delegation hands even the optimization of the boundary to the other party. Each migration presupposes infrastructure: standards standing, context flowing—only then can judgment maturity be tested. Delegation without standards is not trust; it is loss of control. Conversely, with standards and context in place and everyone still pinned at “instruction,” you have the old gene this chapter replaces: execution worship on the talent side—reward only obedience, and judgment atrophies while waiting for orders.

Working a few more years and learning a little more will not carry you from S2 to S3. Between S2 and S3 sits a chasm (Moore’s word, because the structure is isomorphic): the two sides want different things, and kinetic energy piled on the left clears to zero at the edge. Concretely, a fourfold break.

The assessment axis changes: excellence at S2 is executing the standard fast and true; excellence at S3 is making the right trade-off where the standard does not reach. Ten thousand hours of execution practice will not grow one gram of judgment.

Habits reverse: S2’s virtues (do not question the standard; deliver at zero deviation) are exactly S3’s obstacles (question the standard; surface risk). Crossing the chasm means uninstalling some of the habits that made you successful so far.

Feedback structure changes: execution feedback is fast and clear; judgment feedback is slow and foggy, mixed with luck. Without attribution tools, people fall into that vacuum and, after the first misjudgment, retreat to the safe zone.

Responsibility jumps: an S2 error is an execution incident; an S3 error is your judgment wrong—and you sign. Many people are not short of ability; they refuse to sign for outcomes and spend a lifetime on the left bank.

See the four breaks clearly and the ladder’s true identity appears: it is the bridge the organization builds across the S2–S3 chasm. Five rungs are five spans of bridge deck. Each step up adds one unit of judgment responsibility and removes one layer of safety net. At instruction, the manager carries all judgment. At guidance, the member begins to see how judgment is made. At consultation, they judge while both share the load. At authorization, they sign while the boundary still holds. At delegation, even the boundary is handed over. You cannot cross the chasm in one leap, but you can cross it in sections. Organizations cannot wait for employees to grow wings; they must build the bridge. Employees should not expect seniority to slide them across; every span must actually be walked.

Migration acceptance must also be named, or the ladder decays into another seniority escalator. The signal to step up is both: consecutive results at the current level (evidence of judgment quality), plus at least one active surfacing of risk or bad news (evidence of judgment honesty). The second signal is easy to ignore; it tests the hardest of the four breaks—willingness to sign for a judgment. The reverse also holds: demotion is calibration. Someone who twice blows through the boundary at authorization is cheaper—for them and for the organization—returning to consultation and walking a span again than falling a third time in place.

If you stand on the left bank now and the bridge has not yet reached your feet, there is a crossing that does not wait on the organization: write standards for the work you are already executing. Writing a standard forces judgment: what counts as good, where the boundary is, what to do with exceptions—all S3 muscle. Risk is low: write it badly and no one is hurt; write it well and it is seen at once. Second, claim exceptions the standard does not cover. Exceptions are judgment’s natural drills; where others dodge is your bridge deck. Third, use the daily three questions to build an attribution loop, turning each judgment’s right or wrong into accumulable experience. None of the three steps needs a title. The fastest S3 ticket is to write standards for the work you are executing. Judgment is ratified after it has been practiced.

Draw the organization’s responsibility boundary with the same precision. The organization builds the bridge, but the bridge does not carry people on its back. Crossing is, by mechanism, selection: of the four breaks, uninstalling old habits and willingness to sign for outcomes cannot be injected by any external training. So the organization’s whole move is a three-piece set: set standards (write and publish the observable marks of “good” at each level; those observed must know in advance what is observed), open channels (ladder decks plus writing standards, claiming exceptions, and the three questions—open to everyone), build observation (promotion and demotion recognize only behavioral record, not self-report or impression).

Evaluation must be observation, because judgment cannot be measured by talk: interviews yield self-report, reviews are performance, declarations manage other people’s judgments; behavioral record exposes the true level. That is isomorphic to evaluating AI. No one opens a model to confirm it “understands”; you design tasks, observe output, and score against a standard. We already hold the observation instruments (ROI cards, standard-output counts, exception logs, Feynman checks); what is usually missing is the discipline to use them as instruments.

One easy pit sits here: job titles do not carry level information. The same “HR” title: at an ordinary company, S2 executing a recruiting process; at a top company, S4/S5 designing talent standards and observation systems. Same function, two or three judgment levels apart. Managing talent structure by function name is measuring people with an unmarked ruler.

The manager’s role on this ladder is redefined too: management’s end state is more people with reliable judgment rights—remaking yourself from a decision bottleneck into a distributor of decision rights. That is Block’s player-coach.

6. Loser sample: a company that locked decision rights #

When decision rights do not migrate, people die. ofo demonstrated it at a four-billion-dollar valuation.

The mechanism runs deeper than the popular line “Dai Wei was too stubborn.” The verified clause: ofo’s board required all shareholders to sign on financing and other major matters. In the October–November 2017 window for a merger with Mobike, Dai Wei used that mechanism to veto the deal. His words: even if there were a merger, “we would merge Mobike.” He refused even the compromise of becoming sole CEO of the merged company in six months.

Then the clause turned. The same all-sign rule was used in reverse by Didi and Alibaba. For more than a year afterward, no new financing could close; the only thing that could pass was mortgage debt, which needed only Dai Wei’s majority voting rights. The company sank in two-way paralysis: the founder locked the shareholders; the shareholders locked the company.

Users settled the account: by late 2018 more than ten million people queued online for deposit refunds, queue numbers into the tens of millions. That string of digits is the most literal photograph of five words: judgment cannot flow.

The decision-rights ladder was inverted. Clauses designed for defense locked the migration of judgment across the whole organization. Judgment cannot move down (every major judgment inside converges on the founder alone) and cannot move out (the clause blocks any external correction).

The control case is Didi and Kuaidi: before their merger, investors’ veto rights were collectively removed—the same class of clause torn out before a successful combination. Place the two cases together and you get an iron rule for designing decision rights: any clause that stops judgment from flowing, no matter whom it was written to protect, ends by trapping the organization itself.

The R&D unit’s full ledger is already open in section three. Other execution roles are on the same rebuild path with similar effects; their sample is less clean, so it stays closed here.

7. Boundaries of the claim #

Three.

First, collapse is a directional judgment, not a layoff order. Collapse speed differs by an order of magnitude across industries; physically heavy delivery businesses will keep S1/S2 roles for a long time. The message to employees is a clear route across the river: S2 to S3 is a chasm, not a ramp (section five). On the individual side the piers are the four adaptive-insight moves plus the two learning tools of reflection and Feynman; on the organizational side the bridge deck is the decision-rights ladder. For people in the collapse zone, the way out is insight, learning tools, and that ladder. What the organization owes them is to build the bridge.

Second, “everyone a partner” names a structural direction. When every remaining role bears judgment, the relation between organization and member necessarily evolves from hiring labor toward partnering judgment. How equity, distribution, and governance catch up is each company’s own engineering problem. This chapter writes no prescription.

Third, sample quality, stated honestly. Block is early and admits partial breakage may come first—so I cite only mechanism design. ofo’s all-sign mechanism is cross-checked across reports, but parties contradict one another on the veto list—so I cite the mechanism, not the list. The origin of the hiring standard (section four) and the R&D rebuild ledger (section three) are company operating figures, unaudited—hear them at a self-report discount. Two samples are still missing: a traditional enterprise that successfully redrew its talent structure, and a loser that “replaced hierarchical coordination with AI and failed.” The second matters especially. Block is still a single case; a single case can only be a hypothesis.

What to Do Monday Morning (principal-leader view) #

Three steps, inside two hours:

1. Mark S levels: label every role on the org chart with an S level. Do not ask “Can AI replace this person?” Ask “Has the judgment for this role already been finished by someone else?” If yes, it is S1/S2. Count the S3+ share—that is what remains after collapse. Below thirty percent, redrawing the talent structure belongs on the agenda.
2. Mark management level: for each core member, mark current decision-rights rung (instruction / guidance / consultation / authorization / delegation). Then ask: if they cannot step up, is what is missing understanding of standards, sustained results, or the ability to surface risk honestly? The answer decides which lesson you owe them—and whether you have clenched decision rights into a bottleneck.
3. Audit the clauses: open the articles and investment agreements; find every “unanimous consent” or “unilateral veto” clause; ask of each: in the worst case, does this stop judgment from flowing? ofo’s lesson is that such clauses are invisible in ordinary times and unsolvable when they fire.

Note (individual and team view): give yourself an S grade by the same rule—how much of the judgment in your current work has already been finished by others? The higher the share, the more time should move to practicing the four moves. The collapse zone is not terrifying. Standing in it without knowing is.

Chapter Acceptance Self-Check (against the chapter’s five acceptance standards) #

1. Claim restatable in one sentence ✓, and an inference from the core claim (execution → zero → roles whose profession is execution and layers whose profession is hauling information collapse together).
2. Whiteboard framework figure ✓ (form-shift and five-rung decision-rights ladder inserted).
3. External comparison and data ✓: positive Block (verified 2026-07-26: layoff figures, statement text, three-role mechanism; design only, no success/failure verdict); loser ofo (all-sign mechanism verified; Didi–Kuaidi control); R&D rebuild in section 3, hiring-standard origin in section 4 (2026-07-26, company figures, unaudited; quality stated in claim boundaries); two missing samples marked honestly.
4. Twenty-one quotable-line candidates ✓ (v1.4 middle-layer dual ingredients and S3 ticket; v1.5 bridge does not carry people, observational science, job titles).
5. “What to Do Monday Morning” principal-leader three steps + personal note ✓.
6. Fluency ✓: whole-sentence rewriting and English breath under current prose-standard.