Haute Lumière
Commerce · VI.03 · MMXXVI · daylight
TEN CONCEPT BRIEFS · Chapter VI.03 — Holarchy Without Hierarchy
One page each. A reader who reads only these ten pages has the chapter.
BRIEF 1 — The Holon
The idea. A holon is a thing that is a whole and a part at the same time, and both at full strength.
Arthur Koestler coined it in The Ghost in the Machine from the Greek holos, whole, with the particle -on for a part. His image is the Roman god Janus: a holon has two faces, one turned inward at what it contains and one turned outward at what contains it. Inward it is self-assertive — it holds its own rules, its own integrity, its own right to act. Outward it is integrative — it serves the thing above it.
Neither face is optional. A part with no inward face is a cog, and it cannot adapt. A whole with no outward face is a breakaway, and it cannot be coordinated with. Koestler's claim is that every stable complex system — a cell, an organ, a sentence, a firm — is built of holons, and that pathology in such a system is almost always one face swallowing the other.
Worked example. A Buurtzorg nursing team of twelve holds its own intake, rostering, hiring and clinical judgement. That is the inward face, and it is complete: nobody above them approves a roster. Outward, the team carries the brand, the clinical standard and the shared back office of 50 people it does not run. Remove the inward face and it is a district office. Remove the outward face and it is twelve nurses with no employer.
Why it matters. It gives you a diagnostic. When something in an organisation is going wrong structurally, ask which face has been eaten. Micromanagement is the outward face eating the inward one. Empire-building is the inward face eating the outward one. They look like opposite problems and they are the same problem.
You already know this because you have belonged to a family that was genuinely its own thing and also genuinely part of something larger, and you did not experience those two facts as in competition.
BRIEF 2 — Nested Enterprises: Ostrom's Eighth Principle
The idea. A commons of any size holds only if its governance is organised in layers, each layer doing the part of the job that is its own size.
Elinor Ostrom's Governing the Commons set out eight design principles common to common-pool resource institutions that lasted centuries — Swiss alpine pastures, Japanese village forests, Spanish irrigation huertas, Philippine zanjeras. The eighth reads: for resources that are parts of larger systems, appropriation, provision, monitoring, enforcement, conflict resolution and governance activities are organised in multiple layers of nested enterprises.
What the principle actually rules out. Not hierarchy — single-layer governance. A commons run entirely locally cannot handle a dispute that crosses its boundary. A commons run entirely centrally cannot see the local knowledge that makes monitoring cheap. Ostrom's later work called the working arrangement polycentric: many centres of decision, formally independent, functionally interlocked.
The evidence. Cox, Arnold and Villamayor-Tomás reviewed 91 case studies against the principles in 2010 and found them well supported — restating the eight as eleven by subdividing three of them, and discarding none. That is an unusually strong result for a design framework in the social sciences.
Why it matters. The eighth principle is the one most often skipped by people who love the first seven, because it is the one that says you will need more than one level. It is the empirical rebuttal to flatness, offered by the person who did more than anyone to show that ordinary people can govern themselves.
You already know this because you have watched a neighbourhood dispute that the neighbourhood could not settle, and you did not conclude that neighbourhoods were a bad idea.
BRIEF 3 — Near-Decomposability
The idea. Complex systems that survive are ones in which almost all of the interaction happens inside the parts, and only a little crosses between them.
Herbert Simon's "The Architecture of Complexity" is the formal statement, and his parable of the two watchmakers is the intuitive one. Hora builds his watches in stable sub-assemblies of ten pieces; Tempus builds each watch as one thousand pieces in a single sequence. Both are interrupted at the same rate by customers telephoning. Hora loses at most one sub-assembly per interruption; Tempus loses everything since his last start. Hora finishes watches. Tempus goes out of business.
The formal version. A system is nearly decomposable when the within-part interactions are strong and frequent and the between-part interactions are weak and rare. Such systems can be understood one part at a time, evolve one part at a time, and recover one part at a time.
Why this is the load-bearing assumption of the whole chapter. Nesting works because the work is nearly decomposable. Where it is not — a continuously coupled process, a live trading book, a tightly integrated engineering programme — no structure can raise the containment rate, because the interdependence is in the work, not in the org chart.
The test, and it is a real one. Take a hundred decisions from last quarter. Count how many could have been closed correctly by one unit with the information it already had. That fraction is your near-decomposability, and it is measurable in an afternoon.
You already know this because you have refactored something — code, a process, a room — and the whole benefit was that afterwards you could change one part without touching the others.
BRIEF 4 — The Containment Rate
The idea. One number decides whether nesting pays: the share of decisions a unit can close without sending anything upward.
Call it c. It is not a measure of empowerment and not a measure of morale. It is an arithmetic property of a structure that can be counted from a log.
How to measure it. For one quarter, log every decision that required coordination. For each, record whether it was closed inside the unit, or escalated — and if escalated, whether the reason named a genuine cross-domain interaction or a preference. Then:
c = decisions closed inside the unit / all decisions loggedThe two rules that produce different values of c from the same tree. Under an authority-threshold rule, a decision rises when it exceeds a limit — so c is set by the size of the limit and has nothing to do with whether the unit knew the answer. Under a containment rule, a decision rises only when it touches a domain the unit does not hold — so c tracks the actual decomposability of the work. This chapter models the first at 0.55 and the second at 0.85, and both are assumptions, which is precisely why the instrument tells you to measure your own.
Why it matters. Every claim made for self-management — speed, cost, engagement, fidelity — is downstream of c. Argue about c and you are having the real argument. Argue about hierarchy and you are having a different one.
You already know this because you can name, right now, the last decision you were entirely capable of taking and had to send up anyway, and you can remember exactly how long it took to come back.
BRIEF 5 — The Escalation Function
The idea. How far a decision travels is a geometric series, and it has a closed form.
A decision escapes a level with probability 1 − c. So the expected number of levels it traverses in a tree of depth d is:
L(c, d) = ( 1 − (1 − c)^d ) / cThe numbers, at depth 6.
| Rule | c | L | Escalations per decision |
|---|---|---|---|
| Authority threshold | 0.55 | 1.8031 | 0.8031 |
| Containment | 0.85 | 1.1765 | 0.1765 |
| Difference | 0.6266 |
At 240,000 coordination-requiring decisions a year and £270 per escalation level, that difference is worth tens of millions — and it is worth something far more useful than money, which is calendar time.
The property nobody expects. As depth grows, L converges on 1/c. So the gap between the two rules can never exceed 1/0.55 − 1/0.85 = 0.6417 levels, however large the organisation gets. The benefit has a ceiling, and the ceiling is reached early.
Why it matters. It converts an argument about culture into an arithmetic you can do on the back of the meeting agenda — and it tells you, before you restructure anything, the most the restructure could possibly be worth.
You already know this because you have watched a question climb four levels and come back changed, and you knew without being told that each hop cost you both a week and a little bit of the question.
BRIEF 6 — Span and Depth
The idea. Span sets depth, depth sets escalation cost, and both are fixed by arithmetic rather than by preference.
depth = ceil( log_span( units ) ) + 1Worked example. 10,000 people in teams of 12 is 834 teams. At a span of 5 — FEMA's incident-command doctrine gives 1:5 as optimal with a workable range of three to seven — the tree is 6 levels deep and carries 1,043 coordination posts, 10.43% of the front line. Run the same span all the way down to individuals rather than to teams and it is 2,500 posts, 25.0%, which is close to the overhead the Dutch home-care sector actually carried.
What widening the span does. At span 7 the tree is 5 levels and the hierarchy's total coordination cost falls from £130.00m to £123.91m. Widening the span is the hierarchy's own best move.
And here is the sting. A wide span is only safe when the units under it are self-contained, because a manager with nine reports cannot supervise nine dependent processes. Widening the span is nesting, arrived at from the other direction and without the vocabulary.
Why it matters. It means the two structures are not opposed doctrines. They are two positions on one dial, and the dial is containment.
You already know this because the best manager you ever had almost certainly had a great many reports and almost never told you what to do.
BRIEF 7 — The Quadratic Mesh
The idea. Removing the manager does not remove the coordination. It converts it from a hub into a mesh, and a mesh grows as the square.
hub ties = h − 1
mesh ties = h ( h − 1 ) / 2At a team of twelve that is 11 against 66 — 6.0× as many relationships to keep alive. This is the real cost of self-management and it is the one that proponents most often omit.
Where the cap comes from. Set the mesh's maintenance cost equal to a hub's plus a manager's integration time and solve:
h² − 3h − 2(m − 1) = 0At sixty person-hours a month of integration, h* = 12.47 people, with a sensitivity band of 10.46 to 14.16 across plausible integration loads. Buurtzorg caps its teams at twelve. A Morning Star colleague negotiates a CLOU with about ten counterparties. The field found the number before the model did. Across a whole organisation at equal containment the managed hub overtakes the mesh at a team size of 18.82.
Why it matters. A self-managed team that drifts to nineteen is not a flatter organisation. It is a more expensive one, and nobody in it can see why, because the cost is distributed across everybody's calendar rather than sitting in one salary line.
You already know this because you have been in a group chat that was wonderful at six and unusable at twenty, and nothing about the people changed.
BRIEF 8 — The Peer Contract (CLOU)
The idea. Replace the reporting line with a written agreement between the people whose work actually touches.
Morning Star's Colleague Letter of Understanding is the mature form. Each person writes a personal commercial mission, then negotiates directly with the colleagues their work affects: what they will deliver, and by what metrics it will be judged. Gary Hamel's account gives the shape — around 30 activity areas and some 50 metrics, agreed with about 10 colleagues, refreshed annually. Morning Star runs roughly 400 full-time colleagues on about $700m of revenue — $1.75m each — with no managers at all.
Why it is a contract and not a job description. A job description is issued by someone above you and is therefore a statement about authority. A CLOU is negotiated with someone beside you and is therefore a statement about dependency. The second is the thing that actually governs how work gets done, and most organisations have never written it down.
The property that makes it structural rather than cultural. It is readable by a third party. When a dispute arises, there is a document, and the document was agreed by both parties before anyone was annoyed. That is the difference between a norm and an institution.
Why it matters. It is the cheapest single intervention in this chapter. A peer contract can be written inside an existing hierarchy, changes no reporting line, and immediately makes visible the seams that nobody owns.
You already know this because the working relationships you trust most are the ones where you both said out loud, once, what you would do for each other.
BRIEF 9 — The Tyranny of Structurelessness
The idea. A group that declares no structure still has one. It is simply undeclared, and therefore unaccountable.
Jo Freeman's 1972 essay is the canonical statement and nothing since has improved on it. Her argument: structurelessness is impossible; what a group can choose is whether its structure is explicit or implicit. Where it is implicit, influence flows along pre-existing friendship networks, which means it accrues to whoever arrived with standing — and because the power is unacknowledged, it cannot be challenged, voted on, rotated or removed.
Give it a number. A team of twelve has 66 possible ties. The actively-sustained close-tie count is about five per person, which realises 30 ties — 45.5% of the mesh. The other 54.5% still has to carry information, and it carries it through whoever already had the ties.
The measured record. Zappos adopted Holacracy from 2013; in 2015 it offered severance to anyone unwilling to commit and 210 of roughly 1,500 people — 14.0% — accepted. Total turnover that year was about 30% against a historical 20%, an excess of about 150 people. Zappos retired the label in 2020. Medium abandoned Holacracy in 2016 at around fifty people, reporting that codifying responsibilities in detail hindered action.
What cannot honestly be claimed. That exit fell hardest on people without prior standing is the mechanism, not a published figure — Zappos has never broken its attrition out by tenure or demographic. The supporting evidence is adjacent and strong: Castilla and Benard found that organisations describing themselves as meritocratic show more bias in reward allocation, because a declared absence of structure removes what makes discretion checkable.
Why it matters. It tells you what to add back. Explicit delegation, rotation, and information diffused by rule rather than by friendship — Freeman's own remedies, and still the right ones.
You already know this because you have joined a group with no hierarchy and worked out within a fortnight exactly whose approval actually mattered.
BRIEF 10 — The Exception Clause
The idea. A nested structure needs a written, time-boxed, self-reverting way to become a hierarchy for a few days.
Why. An emergency is, definitionally, an event that spans units — that is what makes it an emergency rather than a problem. So an emergency sets the containment rate to zero, and L(0, 6) = 6: the decision climbs the whole tree by consent. At five working days a level that is 30 working days against 5 for a structure with a pre-named commander and unity of command. Twenty- five working days lost. On a same-day consent cadence it is still six against one.
The clause, in four lines. Who may declare. Who commands. The time box in working days. What reverts automatically, on a date, requiring a positive act to extend rather than a positive act to end.
It already exists in working systems. Holacracy's constitution carries one in Article 4.1 — Individual Action — permitting any partner to break the rules in an emergency, with a duty to communicate and restore afterwards. FEMA's incident command system is unity of command with a span of five. Weick and Sutcliffe's high-reliability organisations are hierarchical in normal operation and let decision authority migrate to expertise under stress.
The failure mode, named. The clause is invoked and never rescinded, and the structure becomes a hierarchy that once had a good reason. Automatic reversion by date is the counter, and rehearsing it twice a year is what makes the reversion believable.
Why it matters. Because the honest case against self-management is not that it is slower on average — it is faster on average — but that it is catastrophic in the tail. A clause fixes the tail for the price of a page.
You already know this because you have watched a good crew hand over command to one person during a real emergency without anybody resenting it, and take it back afterwards without anybody having to ask.
All figures in these briefs are computed in lib/verify/VI_03.py and sourced in the chapter's Works Cited. Figures marked REPORTED there are published by the named source; figures marked ASSUMED are the chapter's modelling choices and carry their sensitivity bands.