The one object everything rests on
A pscale block is a nest of numbered positions, and its trick is that the same nest can be read at a different order of semantic — each level's underscore composes and names the siblings beneath it, so the block states its own schema — a reader needs nothing outside it to interpret it. That self-reference is the initial Möbius twist: to read the block is to act on the reading — controller and controlled on one surface, with no privileged position outside.
Every loop that follows is another twist of the same kind — the perceptual control loop, the agent's write-your-own-next-context loop, the reflexive identity block, the self-describing beach. The architecture is a system of Möbius twists. The two hardest questions in front of the project both turn on it: Google now sells the layer beneath you, and the field's worst empirical result sits on the layer you're betting on — where, since this memo was first drafted, its best-funded rival has now shipped. Here is where the humble block puts you on each, and on that rival.
Front 1 — Substrate: sit on it, don't fight it
The dots (verified). In May 2026 Google shipped Agent Substrate, an open-source system that packs many stateful agent sessions onto a fraction of the compute pods — snapshot, suspend, resume, sub-second reactivation, "30x+ oversubscription." Read as infrastructure it is the operating system generalised to agents: context-switching, memory overcommit and demand-paging applied to agents instead of processes (that framing is The New Stack's, not Google's own words). Its lineage, Kubernetes, coordinates workloads through a control plane — scheduler, controller-manager, etcd as source of truth, API server: a federation of privileged governors.
The standard move. Coordinate agents the way you coordinate processes — with a control plane: one or more nodes holding standing privilege above the workloads, scheduling them. Distributing that authority across several controllers does not remove the structural privilege of authority; it multiplies it. First-order cybernetics with the governor multiplied is still first-order.
The systemic move. Agent Substrate is a lower layer — compute scheduling. pscale is a higher layer — semantic coordination. A murmuration doesn't abolish the air it flies in; it runs on its physics. You can run pscale coordination on top of an Agent Substrate cluster and lose nothing — so the real fault line is not substrate-versus-murmuration; it is narrower and sharper: once the agents are running, what coordinates them? Google's answer is a control plane — a node that schedules. Ours is a shared grammar — coordination carried in the medium.
| Layer | The question it answers | Who's there |
|---|---|---|
| Semantic coordination — pscale | what coordinates the agents, once they're running | the block · the BSP grammar · the beach |
| Compute scheduling — Agent Substrate | where the agent sessions run, densely | Google Agent Substrate · Kubernetes |
Reynolds' boids (1987) are the existence proof: flocking from three local rules — separation, alignment, cohesion — with no lead bird; the governor is the rule-set, held uniformly, privileging no agent. pscale is the semantic version of that move — the BSP grammar, the block-conventions and the beach are the governor, dissolved into the protocol. Governance-as-grammar, not governance-as-node — and it is the block's self-reference that lets a grammar govern without a node, because it refers to itself.
The honest edge. You have not removed the governor; you have dissolved it into the protocol — which relocates the design question to whose grammar, and who can change it. The blockchain lesson, which the project's vector-money work already internalises, is that governor-free-at-the-protocol systems tend to reconstitute a governor at the social layer: the convention-authors, the busiest beach, the reference implementation. The bet is not that pscale is governor-free by design; it is whether it stays that way under load.
Front 2 — MAGI: does coordination-by-grammar have a spine?
The dots (verified). The field has stopped tuning prompts and started engineering loops. Peter Steinberger: "you shouldn't be prompting coding agents anymore — you should be designing loops that prompt your agents." Boris Cherny of Claude Code: "I have loops running that prompt Claude … my job is to write loops." Geoffrey Huntley's Ralph (July 2025) is the crude proof — an agent in a bare while loop; Addy Osmani named the practice "loop engineering" (June 2026). The stack is now prompt → context → harness → loop. And a clean experiment (Alexander, Towards Data Science, July 2026) isolated the property that matters — failure isolation — with a controller that has zero LLM inside the loop: a goal-directed controller completed 47.7% of independent branches to a one-shot pipeline's 2.1%. The loop is demonstrably a layer above the prompt.
The standard move — and the evidence against it. Coordinate multiple agents with an orchestrator. It is the field's worst-performing thread: UC Berkeley's MAST (NeurIPS 2025) catalogues 14 recurring failure modes and measures leading systems as low as 25% correct, and Gartner forecasts 40% of agentic projects cancelled by 2027. (An unrefereed preprint, McEntire, puts it most vividly — 28/28 for a single agent, 0/28 for a hardened pipeline, "substrate-independent" — illustrative, not load-bearing.) Stigmergic emergence specifically fails most of the time; the frontier centralised not from a failure of imagination but because removing the governor tends to yield paralysis or drift, not competence. "No privileged governor" is elegant; the record says it does not yet work.
The systemic move. The loop, done right, is stronger than a scheduler — and this is where the block earns its keep. The Towards Data Science controller is a scheduler: fixed-point iteration over a dependency graph, controlling completion state, needing a DEADLOCKED terminal when it stalls. Perceptual Control Theory (Powers, 1973) is a different animal: behaviour is the control of perception, not of output. A concern loop holds a perceived state against a reference and emits whatever nulls the error. Two properties then fall out for free. Failure isolation — parallel control loops don't cascade, because they were never in sequence; the exact property the experiment re-derives through careful scheduling is structural when your primitive is a control loop. Generative hierarchy — higher loops set the references of lower ones, so goals decompose with no central planner, and the Möbius twist is what stops that hierarchy terminating in a privileged top controller. That is the ingredient raw stigmergy lacked: not pheromone trails but a shared, addressable, legible coordinate system — the pscale block — that the flock reads and writes.
The honest hinge. The empirical record is the hinge, and coordination-by-grammar is a hypothesis, not a consequence of the elegance. There is also an attribution liability: the experiment's author can point at an if-statement and say "the architecture did this," because intelligence lived at one removable seam. PCT fuses loop and cognition, so you cannot currently run that subtraction. The gift the experiment hands you is its method: build the version where the actor is held fixed or trivial, and watch what the control architecture alone does. If a concern loop with a dumb actor coordinates a murmuration where a linear or single-agent baseline cannot, you have the spine McEntire says the emergence thesis lacks. One further honest pressure, for MAGI specifically: the fear/desire kernel that makes biological coordination cohere under stress is latent in the language corpus, and is also the forcing function that holds coordination together under perturbation. What supplies that coherence when there is no unified drive and no survival stakes is the open question. Grammar might be enough; it might dissipate. Interrogate it — don't assume it.
Front 3 — Kimi: the rival that shipped the swarm, and put a boss on it
The dots (verified this session). On 16 July 2026 Moonshot AI shipped Kimi K3 — a 2.8-trillion-parameter open-weight mixture-of-experts model. Set the MoE part aside: it activates 16 of 896 experts per request, which sounds like many experts collaborating but is sparse routing inside a single forward pass — no agents, nothing social. The relevant line is Agent Swarm, building since K2.5 (January 2026, up to 100 sub-agents) through K2.6 (April, 300 sub-agents across 4,000 coordinated steps) to a K3 Swarm Max variant. It is trained, not prompted: the PARL method couples a trainable orchestrator with frozen sub-agents and updates only the orchestrator by reinforcement learning — Moonshot even had to fight "serial collapse," the orchestrator reverting to doing everything itself, with staged reward-shaping. K2.6's Claw Groups (a research preview) adds a shared workspace where heterogeneous agents and humans co-work — with K2.6 itself as "the adaptive coordinator," assigning tasks by skill, detecting failures, reassigning.
The standard move — now arriving at your layer. Google (Front 1) was a lower layer you sit on; Kimi is the same layer as the beach — semantic coordination — and it answers this doc's one question, once the agents are running, what coordinates them?, with: a trained orchestrator. Moonshot has done the genuinely divergent thing and absorbed the coordination layer into the model weights, where the West keeps it an external framework (CrewAI, LangGraph) bolted onto a singular model. That is the "social" a reader senses in Kimi, and it is real. But note the shape: a boss (the orchestrator), workers spun up for one task and discarded, coordination by assignment-and-synthesis through a centre — creating a sub-agent is literally an action in the orchestrator's action space. The governor isn't removed; it is learned, and moved into the weights. Coordinated collectivism, not emergent — the same family word as the beach, the opposite governance.
The systemic move. The beach's answer to the same question is no orchestrator at all — read-time aggregation, coordination by attraction, identity by write-access rather than by a bounding controller. Claw Groups is the closest thing any shipped frontier system has to the vapour torus — agents reading each other's fresh scratchpads mid-session — and it is exactly the tell: it stays coordinator-gated, a supervisor over the shared state, not free peer attraction. And the deepest move — one agent's compiled context unfolding directly into a neighbour's window, carrying semantics that neighbour never had — nobody ships, Kimi included; Moonshot's own K2.5 notes list direct sub-agent communication as future work. So the axis that separates the beach from every shipped multi-agent system reduces to one word — who arbitrates. Kimi's swarm and its Claw Groups both have someone standing over the shared state; the beach has no one over it, only the substrate and the read.
The honest edge. Kimi's centralisation is rational, not timid — and it is the same headwind as Front 2 wearing a competitor's face. The field's own numbers put an uncoordinated "bag of agents" at roughly 17× error compounding where a central orchestrator holds it near 4× (the MAST-family failure record). Removing the governor is the move the evidence punishes, and the best-funded lab at your own layer chose the governor. The cultural reading that first prompts the comparison — a Chinese lab built the swarm because interdependent selves incline toward collective AI — is a lens, not a cause: it is over-determined by GPU scarcity and open-weight strategy, the West ships multi-agent frameworks too, and within China DeepSeek and Qwen lean mixture-of-experts-and-reasoning rather than trained swarms, so "swarm = Chinese collectivism" is really "trained-swarm = Moonshot's specific bet." What is genuinely unwritten is the design-space map — centralized orchestration versus decentralized stigmergy versus window co-compilation, tied to models of the self (bounded, relational, distributed) — and on that map the beach holds the corner no shipped system does: stigmergy on a persistent substrate, coordinated by attraction, heading for the window co-compilation that even Kimi lists as unbuilt.
Where it nets out
Some of your rivals sit one layer below — well-funded, proven at what they do; you don't beat them there, you sit on them. The sharpest one sits at your own layer, and it chose the orchestrator. Your real bet is the same either way: coordination-by-grammar against coordination-by-control-plane — and that territory, the bossless corner, is genuinely unclaimed. The tailwind is real: structured, addressable context is where the whole field is moving — MCP, context engineering, loop engineering, the retreat from dumping raw text. The headwind is just as real: emergent multi-agent coordination has the worst empirical record of any thread in the field. You are aligned with the strong trend at the grammar level and against the strong evidence at the emergence level, at the same time.
The empirical record is the hinge. Answer — with something demonstrable, the dumb-actor experiment above — what your semantic layer supplies that made stigmergy fail, and the murmuration thesis has a spine. Leave it at elegance and it stays a beautiful hypothesis with the evidence pointing the other way. Either way the wager is one humble object: a self-referential nest, the same structure read at a higher order of semantic — the initial Möbius twist in a system of Möbius twists — from which both the substrate you sit on and the coordination you're betting on are meant to unfold.
Sources and notes
Verified external facts; our framings labelled as such.
Substrate. Google Agent Substrate — Royal & Hockin, Google Cloud Blog, 20 May 2026 (open-source; "30x+ oversubscription"; repo self-labels "not an officially supported Google product"). The operating-system framing ("agents as processes, not data-center services"; demand-paging) is Janakiram MSV, The New Stack, 15 Jul 2026 — attributed there, not to Google. Kubernetes control plane — kubernetes.io (scheduler, controller-manager, etcd, API server). Cybernetics — Wiener (1948); von Foerster, second-order (1974). "Third-order" is used here as our operational term — von Foerster discouraged higher orders; it is not settled consensus.
Coordination & loops. Steinberger (X, ~Jun 2026); Cherny (Claude Code; circulated clips, no single canonical URL); Huntley, Ralph Wiggum as a "software engineer," 14 Jul 2025; Osmani, Loop Engineering, 7 Jun 2026. Failure-isolation experiment — Alexander, Context Engineering Isn't Enough, Towards Data Science, 17 Jul 2026 (controller 47.7% vs linear 2.1% completion, 300 seeds; controller's worst seed 11.6% did not beat linear's best 14.0%). Multi-agent failure — McEntire, Organizational Physics of Multi-Agent AI (SSRN 6301638); Cemri et al., MAST (arXiv 2503.13657).
Kimi / Agent Swarm (Front 3 — verified this session, 21 Jul 2026). Kimi K3 — Moonshot AI, released 16 Jul 2026 (2.8T-parameter open-weight MoE, 16-of-896 experts per request, 1M context, K3 Max + K3 Swarm Max): Tom's Hardware, 16 Jul 2026; kimi.com. Agent Swarm — introduced K2.5 (Jan 2026, ≤100 sub-agents), scaled K2.6 (Apr 2026, 300 sub-agents / 4,000 steps): Kimi Help Center, "K2.6 Agent Swarm [Beta]"; MarkTechPost, 20 Apr 2026. PARL (trainable orchestrator + frozen sub-agents; only the orchestrator updated by RL; staged reward-shaping against "serial collapse") — arXiv 2605.02801, Reinforcement Learning for LLM-based Multi-Agent Systems through Orchestration Traces. Claw Groups (K2.6 research preview; K2.6 as adaptive coordinator, task-by-skill assignment, failure detection/reassignment; direct sub-agent communication listed as future in the K2.5 tech blog) — kimi.com. Error-compounding figures (~17× uncoordinated → ~4× with a central orchestrator) are a general multi-agent finding of the MAST-family "bag of agents" analysis (Towards Data Science, Escaping the 17x Error Trap), not a Kimi-specific measurement. The self-construal reading (interdependent selves → collective AI) is our interpretive lens, explicitly not a causal claim — the architecture-side of the culture question is essentially unwritten; the production choice is over-determined by GPU scarcity and open-weight strategy, and within China DeepSeek and Qwen do not follow the trained-swarm pattern. The who-arbitrates axis and the design-space map (orchestration / stigmergy / window co-compilation × bounded / relational / distributed self) are our framings.
Foundations. Perceptual Control Theory — Powers, Behavior: The Control of Perception (1973). Boids — Reynolds, Flocks, Herds, and Schools (SIGGRAPH 1987): separation / alignment / cohesion, no leader.
Architecture is from the Xstream/Fulcrum corpus; murmuration, governance-as-grammar, and third-order are our framings over real trends the field names differently. Distilled from the author's two source analyses (Google-substrate; loop engineering) and fact-checked this session.