The Observer Block Forgot to Build

Jack Dorsey and Roelof Botha published "From Hierarchy to Intelligence" in March. It argues that corporate hierarchy is a two-thousand-year-old information routing protocol, that AI can now do the routing, and that Block is rebuilding itself accordingly: a company world model at the center, three human roles at the edge, no permanent middle layer. The essay landed weeks after Block cut 40% of its workforce, which colors the reception but not the argument, and the argument deserves engagement on its merits.

The history is one of the best short account of org design I have read. Roman Empire span of control, the Prussian General Staff, McCallum's railroad chart, Taylor, the matrix. The diagnosis is right: managers exist mostly to aggregate context upward, relay decisions downward, and hold alignment sideways, and a system that reads every artifact a remote-first company produces can do that continuously.

The critiques so far attack the right essay in the wrong place. Meg Bear points out that DRIs are named without real authority. Carlo Torniai questions whether the model can match the accumulated judgment managers carry. Both are per-decision arguments: who decides, and how well. The structural problem is different. Hierarchy did a second job, silently, and the essay never names it.

Two jobs, one replaced

Routing is the job everyone can see. The second job is carrying the integral.

A middle manager knows how many times this team has been redirected this quarter. How much appetite for change remains. How many "temporary" arrangements have quietly become permanent. Which reversible decisions are no longer reversible because the person who could reverse them left. None of this is a decision. It is accumulated state, and someone was reading it.

Block's design evaluates decisions one at a time. The world model surfaces context, a DRI or IC makes the call, the system re-optimizes, repeat. Every individual decision can pass every individual check. Here is the mechanism that design cannot see:

A one-way door taken in installments passes every per-decision review.

Bezos's two-way doors work as a classification only if door type is observable per decision. It is not. Reversibility is a property of the series. Reprioritize a team once, reversible. Do it six times and your senior engineers walk, which is not. Each extension of a "temporary" workaround is defensible. The workaround becoming load-bearing infrastructure is not a decision anyone made. The dangerous case is never the misclassified one-way door, which better review catches. It is the one-way door assembled from parts, each of which was genuinely two-way at the moment it was evaluated.

No per-decision gate can catch this, even in principle, because the failure is not in any decision. It lives in the integral over the decision history, and the integral is not an input to any single evaluation. This is not a claim about AI capability. A smarter gate does not help. The blindness is architectural.

We have seen this movie

2008. Every mortgage passed underwriting. Every tranche passed its rating. Every per-decision gate was green. The failure was correlated leverage accumulating across a series of individually defensible positions, and the institutions that survived were the ones with a risk function that did not report to the trading desk. The ones that died had made the optimizer its own auditor.

Once you see the pattern, it is everywhere. Term limits do not ask whether the incumbent is performing well; they fire on accumulated tenure regardless of per-decision quality. Margin calls, capital ratios, audits, statutes of limitation, sabbaticals: every one is an accumulation monitor. Every one fires on state rather than decisions. Every one was invented by trial and error after an accumulation failure, and every one is resented in the moment it fires, because in that moment no individual decision is wrong.

These institutions are humanity's answer to a problem we never stated formally: the thing doing continuous optimization cannot also be the thing that notices it has optimized itself through a one-way door. The observer has to sit outside the loop.

What Block should build

Not managers. The essay is right that the routing layer is replaceable. The fix is an instrument:

Let the intelligence layer re-optimize freely inside certified reversible space. Run a separate monitor that tracks accumulated commitment across the decision series: what it would cost, today, to return to the state before this series began. When that integral crosses a threshold, freeze and escalate to a human. Not per-decision veto, which recreates the bottleneck. A tripwire. Rare enough that standing arbitration becomes on-call arbitration, which is the flat-organization economics Block wants, made safe.

The escalation target stays human for a reason the essay gestures at but does not follow through. When the tripwire fires, someone loses, and the loser accepts the ruling because a person owns it. Judges are not information routers. Their product is legitimacy, and legitimacy does not compile.

The prediction

If Block ships the optimizer without the observer, it will fail in a specific, recognizable way. Not a bad decision anyone can point to. Thrash: priorities that re-optimize continuously, commitments nobody can build on, a decision series optimal at every step and ruinous in aggregate. The postmortem will find no wrong call, because there was not one. The wrongness lived in the integral, and nobody was assigned to read it.

Dorsey and Botha closed by welcoming pressure testing. This is mine: you are not removing hierarchy, you are decomposing it. Routing to the model, line command to DRIs, and the third function, the one hierarchy did silently and badly, to an instrument that does not exist yet. Build the observer, or rediscover why every durable institution already has one.

---

References

The essay: Jack Dorsey and Roelof Botha, From Hierarchy to Intelligence, Block, March 31, 2026.

Responses worth reading, and where this post differs:

• Meg Bear, Hot Take on Jack Dorsey's "From Hierarchy to Intelligence" Manifesto. DRIs named without real authority. A per-decision authority critique; this post argues the gap is upstream of authority.
• Carlo Torniai, Intelligence Over Hierarchy: Block's Bet and Its Blind Spots. Physical-world knowledge, labor law, and whether the model can match accumulated managerial judgment. A judgment-quality critique; this post argues no per-decision judgment, however good, sees the integral.
• From Hierarchy to Intelligence: A Reflection from the Trenches of AI in Biotech. The world model is overpromised and change management decides the outcome. An execution critique; this post argues the design has a structural gap even if execution is flawless.
• Moore Dagogo-Hart, From Hierarchy to Intelligence: A Different Path. Same diagnosis, different prescription on the human side.

Context: Fortune's coverage of the essay and the February workforce reduction.