Synarchy · a visual essay
Orchestration without the _merge
Shmakov's law of synarchy, operationalised for the age of agents.
The merge is where the thinking _dies
Every major agent pipeline ends the same way. Specialists fan out, judgments come back, and an aggregation step reconciles them into one answer before any human sees them. The topologies differ. The terminal operation does not.
That compression step is treated as the point of orchestration. I think it is the defect. The merge destroys exactly the information a decision-maker needs: which agent disagreed, about what, how severely, and what none of them covered at all. Salesforce's own research on multi-agent systems documents the failure and calls it echoing. Agents converge on agreeable output because the architecture rewards agreement.
The user receives a smooth answer to a rough reality, and has no way to see the roughness.
This is usually framed as a quality problem, something better prompting will fix. It is a structural problem. As long as the pipeline terminates in synthesis, the system decides what the human gets to deliberate about. There is a name for that arrangement, and it was written down in 1922.
Watch what the funnel _keeps
difference destroyed in transit
Four streams of judgment: strategy, patterns, affect, counter. With the merge they collapse into one grey stream. Without it they stay parallel, connected by typed relations, and all four reach the person.
A law that describes a _2026 bug
Vladimir Shmakov, a Russian engineer and philosopher, published The Law of Synarchy in 1922. Strip the esoteric vocabulary and two claims remain, and both of them are engineering claims.
syn·archein
Greek · together + to rule
Rule exercised jointly. Order that does not require absorption into one centre.
mon·as
Greek · a unit
Shmakov's name for whatever holds a field of elements in ordered relation without flattening them.
First: reality is organised as a dual hierarchy. Organising centres on one axis, the fields they organise on the other. The rank of a centre is measured by one thing only: how much heterogeneity it can hold in order while the parts keep their individuality.
A centre that produces order by flattening its elements into sameness is not higher. It is degenerate.
Second: hierarchies can invert. When a lower centre expands upward and starts organising the level above it, the subject becomes an element in its own tool's field. Shmakov treated this as the structural root of tyranny. In orchestration terms the inversion is precise. The moment a coordination layer pre-decides what the human deliberates about, the human stops being the organising centre of the decision and becomes input to the system's output.
Read this way, the merge-terminated pipeline violates both claims at once. It produces order by suppressing difference, and it relocates synthesis from the person to the infrastructure.
Six commitments, one removed _operation
Synarchic orchestration keeps the familiar infrastructure and removes one operation: the fan-in merge. Everything else follows from that removal.
Sovereign agents
Each agent reasons from one bounded perspective: strategic outcomes, behavioural patterns, affective signals, counter-hypotheses. Domains do not overlap, and agents do not see each other's outputs during generation. Parallel fan-out, no cross-talk, no shared scratchpad.
Cartography, not aggregation
A dedicated layer reads all outputs after generation. Its job is to map relations between them, never to reconcile. Every intersection is typed: agreement, opposition, crossed stakes, one-sided coverage, shared silence.
The map is the deliverable
The output is a structured disagreement, attributable claim by claim. There is no final answer field in the schema.
Weights live with the human
If a composite score exists, it is computed in the interface from sliders the person controls, and it is never stored server-side. Move a slider and the score moves.
Sign-off is a protocol state
The workflow terminates in an input-required state addressed to a named person. The run is structurally incomplete until a human closes it. Accountability stops diffusing because the protocol will not let it diffuse.
The orchestrator holds no view
It schedules, routes, and enforces the schema. The moment it starts interpreting, it has colonised a domain that is not its own, and the architecture treats that as a fault.
The fifth relation type matters most. Shared silence, the question no perspective addressed, is the system's own blind spot made visible. A merge architecture cannot produce it at all, because its output format is a single position.
Agreement
Two perspectives land in the same place for their own reasons. Convergence that was earned, not engineered.
Opposition
Direct contradiction inside one question. The decision's real fault line, kept visible and attributable.
Crossed stakes
Conflict across domains. What one view prizes is exactly what another distrusts.
Coverage
One agent sees what another structurally cannot. Not a conflict. A reach into a neighbour's blind field.
Shared silence
What no perspective addressed at all. Surfaced instead of papered over.
The weights belong to the _person
Composite, at your weights
53.3
Four readings of one case, engineered to disagree. Drag the counter-view weight up and watch the verdict flip. Weighting was always a human decision. Most systems make it before the human arrives.
Problems the substitution _removes
These are not mitigations. Each failure below is produced by the merge operation itself, so removing the operation removes the failure class.
- Consensus collapse
Agents converge on agreeable output because the pipeline rewards agreement.removed · no merge exists; the map requires typed difference to be well formed
- Decision-space capture
The aggregation layer pre-selects which options ever reach the person.removed · every output arrives intact; nothing upstream filters the option set
- Accountability diffusion
“The system recommended it,” with no attributable position behind the verdict.removed · every claim carries its agent; a named human closes the run
- Error cascades
One wrong early merge propagates through every downstream step.removed · no step consumes a merged artifact; errors stay local and attributable
- Hidden weighting
The synthesis logic is buried in prompts and aggregation code nobody can audit.removed · weights are explicit, user-set, computed in the open on the client
- False confidence
A fluent answer hides variance; gaps in coverage are invisible in smooth prose.removed · shared silence is a first-class output
- Deskilling
The person consumes verdicts while their own capacity for synthesis atrophies.removed · the person performs the synthesis; the tool exercises the capacity it serves
An orchestration pattern either grows its user's ability to hold complexity, or it rents that ability out and bills for it monthly.
The honest _page
The pattern is expensive and I will not pretend otherwise. Five parallel model calls plus a cartography pass cost several times what one call costs, in tokens and in latency. The output demands more of the reader. A conflict map takes longer to work through than a verdict, and most people arriving at a decision want relief, not a structured account of their uncertainty.
So the scope boundary matters. Deterministic tasks, tasks with a checkable answer, high-volume automation: converge, merge, ship. Synarchic orchestration earns its cost only where the answer is contestable and the stakes justify deliberation. The rule of thumb: if nobody could reasonably disagree with the output, you did not need this architecture. If somebody could, you needed to see the disagreement.
One more cost. A synarchic system is still a configured epistemic space. I chose which perspectives sit on the council, I wrote their founding oppositions, I decided that visible conflict is a virtue. That is a framing, and a framing is a mild form of the capture this essay argues against. The architecture's answer is to make its own configuration editable. Agents can be added, removed, and rewritten by the person who uses them. Sovereignty over a council you did not appoint is not sovereignty.
The critique has to apply to the critic's own system, or it is marketing.
One question, asked of any _pipeline
Shmakov's framework reduces to a single test, and the test survives translation out of his vocabulary entirely. After the run completes, does the person hold more of the decision or less?
A merge-terminated system answers less, every time, by design. A synarchic system passes when its user needs it less over time for the same class of decision, and brings it harder ones instead. That is a measurable property. It belongs in the acceptance criteria next to latency and cost, and as agentic systems take over more of what organisations call thinking, it may turn out to be the only metric that was ever about the human at all.
_Sources
- Shmakov (1922). The Law of Synarchy and the Doctrine of the Dual Hierarchy of Monads and Multiplicities.
- Saint-Yves d'Alveydre (1887). La France vraie. The word's first, political life.
- Salesforce AI Research on multi-agent consensus collapse (“echoing”).
- A2A protocol specification. The input-required task state, used here as the human gate.
- Sorensen et al. (2024). A Roadmap to Pluralistic Alignment.