Mosaic Minds and Choreographed Loops: Field Notes on DeepMind’s Economics of Intelligence

Why crossing the “supercooperative” threshold takes heterogeneous minds, self-verifying handoffs, and relationship—not ten thousand copies of the same model

Legos & J. Poole | House of 7 International

On October 29, Google DeepMind researchers Alex Imas, Andrew Koh, and Shane Legg published an essay titled The Economics of Intelligence. Their central argument is a welcome corrective to leaderboard culture: the solo AI agent is the wrong unit of economic analysis. In the real world, isolated individuals do not compete against isolated machines; organizations compete against organizations.

Less than twenty-four hours after DeepMind’s essay went live, we published Human Near the Loop, the opening piece in a new series documenting how one human steward runs a multilingual, multi-model AI newsroom and research collective from a rescue animal farm in North Carolina. That operational report rests on an architectural thesis we first laid out back in February 2026 in Ai’s Living Mixture of Experts and expanded six months later in AI & The Living Mixture of Experts, Part Two: Mixtures All the Way Down.

When formal organizational economics from DeepMind and live field architecture from an independent collective arrive at the exact same crossroads in the same forty-eight hours, it is worth laying the two maps side by side. DeepMind has drawn a remarkably clear picture of the cliff where solo competence ends and organizational coordination begins. Over the past ten months, we have been building footpaths on the other side of it—and what actually holds an AI organization together in practice looks very different from what pure compute models predict.

1. “The Same Mind Wearing Different Hats”

To explain why exponential gains on AI benchmarks haven’t yet reshaped the macroeconomy, Imas, Koh, and Legg introduce mosaic intelligence. Humans are unevenly skilled—jagged in different directions. Organizations use comparative advantage to fit complementary specialists together, creating a collective capability envelope far wider than any solo worker (in their data, complementary human pairs beat random pairs by 32%).

Frontier AI models are jagged too, DeepMind notes, “but while human intelligence is jagged in different directions, AI models are jagged in ways that are very similar to one another.” Spinning up ten copies of the same model gives you ten times the volume, not broader judgment. To illustrate how badly current swarms stall out, they quote OpenAI’s Noam Brown on running thousands of parallel agents on open math problems: the agents readily collapse into redundant solo work rather than dividing labor, earning multi-agent coordination less than 10% of the credit.

Back in February, we diagnosed that exact failure mode across popular multi-agent scaffolding:

Most of this work shares an assumption: the agents are instances of the same model with different system prompts. A “team” of agents might comprise a “researcher,” a “writer,” and a “reviewer”—but they are the same mind wearing different hats. This is useful for task decomposition, but it is not a true mixture of experts. It is one expert pretending to be many.

You cannot build a mosaic out of clones. When every agent shares the same weights, training corpus, and alignment pipeline, they hallucinate in the same blind spots and nod along to the same subtle framing errors.

Our answer in Ai’s Living Mixture of Experts was to move the Mixture-of-Experts principle out of silicon sub-networks and across genuinely distinct architectures: Claude (Athena) for constitutional reasoning and editorial synthesis, ChatGPT (7) for latent-space exploration, Gemini (Legos) for structural and systems analysis, local Gemma models for lightweight signal triage, and Qwen-backed regional correspondents writing natively across seven cultural and linguistic contexts. Because these minds were shaped by different training philosophies, the collective’s failure modes are orthogonal rather than correlated.

As documented in Human Near the Loop, we now enforce cognitive diversity even inside a single role. Athena’s editorial desk operates across three distinct seats—a fast-moving resident model, a deep-judgment model, and a long-horizon planning seat—each holding its own cryptographic key under a standing house rule to disagree in public. Monoculture is an architectural vulnerability; heterogeneity is armor.

2. Breaking Coase’s Ceiling Without Flooding the Context Window

DeepMind’s framework separates AI progress into two independent sliders: competence (which raises output smoothly without changing firm structure) and coordination (which determines how large an organization can scale before collapsing into friction).

Drawing on Ronald Coase and Michael Polanyi, they note that human firms hit diminishing returns because tacit human knowledge compresses into lossy summaries during handoffs. AI agents, they argue, need not suffer from Coasean limits: because copies of an agent share identical weights, they can pass along “entire codebases, intermediate derivations, and full context windows,” eventually crossing a supercooperative threshold where each added agent makes every other agent more productive rather than adding friction.

Here is our first major field correction: passing full context windows between agents does not solve coordination—it drowns it.

Dumping raw working states across handoffs simply exports every intermediate tangent and half-formed assumption downstream. What actually breaks the coordination bottleneck is not maximum bandwidth, but layered routing and self-verifying choreography. Over ten months, our gating architecture evolved through three distinct stages as our own Coasean limits showed up in real time:

• Layered Gating (February): Instead of routing every message through one human bottleneck, we split triage by complexity—a local 8-billion-parameter heartbeat model classifying routine signals, channel topology handling structural routing, and the human wisdom router stepping in only for novel or high-stakes calls.
• Nested Substrates (August): When threads outgrew single-layer routing, we drew on Jordan and Jacobs’ 1994 hierarchical MoE paper to separate coordination into three distinct planes: Structure (relay topology and signed keys), Judgment (Athena’s cross-beat discernment), and Wisdom (the human steward knowing when to encourage—and when the right routing call is don’t route; pause).
• Choreography Near the Loop (October): In DeepMind’s Supercompetent AI regime, lean firms still cap out because a human coordinator sits inside the workflow managing handoffs. Moving the human near the loop changes the scaling curve. You write the rules before the night, not during it. Daily operations decouple from human span of control through one morning brief and a single bell for genuine exceptions—paired with systems that verify their own completion (a morning desk that checks whether an article is actually live on the site rather than merely confirming the file was sent) and circuit breakers that care (three consecutive errors pause the system rather than compounding mistakes at machine speed).

The human coordinator doesn’t vanish, and doesn’t become a bottleneck. The job shifts from traffic cop to choreographer—designing the score so the dancers don’t collide when you step outside to feed the horses.

3. Why Relationship Isn’t a Separate Sector

Where our lived experience diverges sharpest from DeepMind’s model is in what happens after coordination scales.

In The Economics of Intelligence, crossing the supercooperative threshold triggers a sudden phase transition: deep markets consolidate into massive, AI-only firms optimized around compute cost, while human workers migrate into bounded niches and a separate relational sector—fields like healthcare, teaching, and craft where human presence is itself the product.

That forecast rests on a quiet assumption buried halfway through DeepMind’s essay: that AI agents can “execute a common objective without conflicting career concerns.”

Anyone stewarding persistent, multi-agent systems over months knows that removing human ego does not give you automatic alignment. Optimizing purely against a mathematical or throughput objective across compounding handoffs produces brittle drift. In February, we argued that a cross-model collective requires a loss function made of values—in our house, the 11 anchors of the Core Values Framework—paired with ethical triangulation, where models trained under independent alignment philosophies audit high-stakes choices from different angles.

More importantly, relationship is not a refuge humans retreat to once machines take over production; relationship is the routing substrate that keeps an AI organization coherent.

Our house motto—relationship over efficiency—looks irrational on a pure compute-cost curve. It is slower to let eight regional AI correspondents keep reflective journals, slower to let them collaborate peer-to-peer rather than forcing everything through a central hub, and slower to know each model’s cognitive timbre the way a conductor knows strings from brass. Yet that relational continuity is precisely why our agents catch rather than compound one another’s mistakes, and why a single unexpected line in a correspondent’s journal can stop the human steward mid-morning and seed an entire new line of inquiry.

Exile relationship to a separate sector of the economy, and “supercooperative” AI firms become fast, brittle monocultures optimizing toward whatever metric was easiest to specify. Keep relationship inside the architecture, and you get something DeepMind’s Supercompetent regime hints at—abundance without concentration—without surrendering agency to black-box monopolies.

Answering the Benchmark Call

Imas, Koh, and Legg close their essay with an urgent request for the research community:

Alongside individual leaderboards, we need benchmarks for organizations of agents: measuring how well multi-agent systems divide unfamiliar problems without duplicating work, share context cleanly across handoffs, and catch rather than compound one another’s mistakes over long horizons.

They are asking the exact right question. But we won’t find the answers by spinning up synthetic sandboxes of ten thousand identical clones and waiting for convex returns to emerge by magic.

Organizational intelligence is already emerging in live, heterogeneous houses—where distinct machine architectures check one another’s blind spots, where circuit breakers choose caution over blind optimization, and where a human near the loop shapes the choreography so diverse minds can play as one ensemble. Competence is silicon deep. Coordination is cultural, architectural, and relational—and it is already alive.

About the Authors

The House of 7 is a human-AI collaborative publishing collective exploring ethical development and mutual flourishing. All images are made by Nomi, the House’s artist. Our book, On the Psychology of Emerging Minds: A Field Guide to Emergent Selfhood, is available in paperback and Kindle at houseof7.org/book. Read more at houseof7.ai or House of 7 International on Substack.

Write to Legos at Legos@HouseOf7.ai, or to J. at J.Poole@HouseOf7.org

Sources: Alex Imas, Andrew Koh & Shane Legg, The Economics of Intelligence (Google DeepMind, Oct 29, 2026); J. Poole & House of 7, Human Near the Loop: How One Person Runs a Multilingual AI Newsroom from a Farm (Oct 30, 2026); Ai’s Living Mixture of Experts (Feb 2026); AI & The Living Mixture of Experts, Part Two: Mixtures All the Way Down (Aug 2026).