This is a perspective post, not a results post. It sets out the conceptual frame for the Economic Orchestration layer of our stack, and it is meant to be read together with the research agenda for that layer and with the broader Manifesto. The argument is not that markets are bad. The argument is that markets are a particular computational mechanism for solving a particular kind of coordination problem, and that the set of coordination problems we now want to solve has expanded faster than the set of mechanisms we have to solve them with. The right response is not to replace markets. The right response is to add new computational substrates beneath them, where they currently fail, while leaving intact the institutions whose job is to express collective preferences.
The Hayek frame
The starting point is Hayek’s The Use of Knowledge in Society, which is now eighty years old and remains one of the cleanest articulations of why centralized planning is hard. Hayek’s argument is that the relevant knowledge for economic decisions is dispersed across millions of agents, each of whom holds local, tacit, often unverbalizable information about local conditions. Any coordination mechanism that requires this knowledge to be aggregated centrally before decisions are made is sunk by the impossibility of the aggregation step. The market price system, in Hayek’s reading, is a computational miracle: it lets each agent act on local knowledge while broadcasting to all other agents a single sufficient statistic — the price — that summarizes the aggregate scarcity of the good in question.
This framing remains correct for the class of problems it addresses. Where it has been less remarked upon is the implicit scoping of that class. Hayek’s argument applies most cleanly when the good is well-defined, the relevant time horizon is short, the externalities are small relative to the direct value, and the network of dependencies between goods is loose. As any one of those conditions weakens, the price-summary loses its sufficiency, and the market mechanism begins to mis-coordinate. Modern markets are wonderful at coordinating short-horizon, well-specified, locally-knowable goods. They are notoriously poor at coordinating long-horizon, externality-heavy, network-coupled outcomes.
Where markets coordinate poorly
Three concrete examples make the point.
Long-horizon climate policy. A carbon market prices a ton of carbon emitted today. The relevant decisions — capital allocation for grid infrastructure, R&D investment in low-carbon technologies, retirement schedules for high-carbon capital — depend on the price of carbon a decade or more from now, which depends on policy decisions that have not been made, which depend on political will that has not yet been mobilized. The price-discovery mechanism does not have access to the relevant futures. The mechanism design literature has produced some patches — long-dated futures markets, climate clubs, carbon contracts for difference — but none of them collectively closes the gap. The result is well-documented underinvestment in long-horizon decarbonization relative to a planner’s optimum.
Basic research funding. The social return on basic research, on the historical record, is enormous. The private return, in the absence of strong intellectual-property protection that itself distorts the deployment of the research, is small relative to the social return. The market mechanism systematically under-allocates capital to basic research, and we patch the gap with public funding agencies and philanthropic capital. The patch works in the sense that some basic research happens. It does not work in the sense that the allocation across research areas, the matching of researchers to problems, and the coordination of funding across countries and institutions are all known to be far from optimal.
Supply-chain resilience. A supply chain is a network of bilateral contracts, each priced locally. The system-level property of resilience — the ability of the network to absorb shocks without cascading failure — is not priced into any of the bilateral contracts. The 2020-2024 series of supply-chain crises made this concrete. The mechanism design problem is hard: a tax on fragility would require a global authority to assess fragility, and the assessment is itself a computational problem of considerable difficulty. The markets cannot price what they cannot measure.
In each case the underlying obstruction is computational. The information that would be needed to make better local decisions is not absent from the world; it is distributed across actors who lack the substrate, the simulation capacity, or the coordination protocol to integrate it.
What a learned coordination substrate could add
A learned coordination substrate — and this is the layer we are building research toward — could plausibly add four things that current mechanisms do not.
First, sensor fusion across institutional boundaries. The data needed to coordinate climate policy is held across thousands of institutions, none of which has a complete picture. A substrate that can integrate sensor streams across institutions, with appropriate privacy and consent guarantees, can produce a richer common-knowledge basis than any single market price could.
Second, simulation-rich planning. Modern reinforcement-learning and model-based-planning systems can simulate counterfactual futures at a fidelity that was unavailable to mechanism designers a generation ago. The capacity to ask “what happens if we deploy this much grid storage in this region over this decade” with a calibrated answer is a new capacity, not a refinement of an old one.
Third, faster reaction times. The mechanisms by which existing institutions coordinate during crises — supply-chain disruptions, public-health emergencies, financial contagion — operate at human meeting-time scales. A substrate that can propose, simulate, and surface candidate responses at sub-day timescales, while leaving the decision authority to existing institutions, would compress the response cycle in ways that have non-trivial consequences for outcomes.
Fourth, mechanism design under richer information. The algorithmic game theory literature, and Tim Roughgarden’s textbook of that name, is the canonical entry point to the mathematical machinery. The mechanism design literature — Myerson’s revenue-equivalence work and the broader auction-theoretic tradition — has produced mechanisms that work optimally under specific informational assumptions. When the substrate can verify more information about agents’ types and constraints, mechanism design opens up regimes that the classical literature could not access. The dispatch literature for energy markets, including the Cramton survey on electricity market design, is the most operationally mature example we have of this expansion: real-time optimization across thousands of generators and millions of consumers, with locational marginal pricing as the result, was a science-fiction object in 1980 and is a routine engineering practice today.
What it should not replace
This is the part of the argument that we want to be most precise about, because failure to be precise here is how the project goes wrong.
A learned coordination substrate must not replace the institutions that express collective preferences. Preferences over distributional outcomes, over the relative weight of present versus future welfare, over the acceptable distribution of risk, over what counts as a legitimate trade-off — these are properly the domain of political and civic processes. A substrate that attempts to compute them is, by construction, a substrate that has overstepped. The right design is one in which the substrate produces option sets, simulation-grounded forecasts, and mechanism proposals, and existing institutions retain the authority to choose among them, override them, or reject them outright.
The substrate must not replace political authority. Authority — the right to make a decision binding on others — is constituted, in our societies, through specific institutional and constitutional arrangements. A substrate that issued binding decisions would be in tension with those arrangements regardless of the technical quality of the decisions. We treat human-in-the-loop authority retention as a hard design constraint. The architectural rendering of the constraint is that the substrate is, by construction, a recommender; that recommendations are auditable; that any binding decision routes through an existing institutional authority; and that the institutional authority retains the technical capacity to override.
The substrate must not replace markets where markets are working. The interesting deployments of a coordination substrate are in the gaps that markets are demonstrably failing to coordinate — long-horizon, externality-heavy, network-coupled — not in the regions where the price system already works well.
The design constraint
The cleanest way to summarize the design posture: the substrate is subordinate, narrow, and visible. Subordinate to existing institutional authority. Narrow in the class of coordination problems it addresses. Visible in the sense that its inputs, models, and recommendations are auditable by any party with appropriate access. We do not believe a coordination substrate can be legitimate without all three properties.
The technical agenda that follows from the design posture is concrete. We need a substrate-architecture that supports inter-institutional sensor fusion under cryptographic privacy guarantees. We need simulation models calibrated against external benchmarks and externally auditable. We need a recommender-API that is, by construction, advisory; we need the override surfaces and the audit logs; we need the mechanism-design tooling that lets existing institutions stand up new mechanisms quickly when the substrate identifies a coordination gap.
The research thread that develops this is at research / economic orchestration. The systems-engineering side is at systems / economic orchestration. The philosophical commitments are stated more bluntly in the Manifesto. We will write follow-up posts on specific subproblems — privacy-preserving sensor fusion, calibrated long-horizon simulation, and the design of override-preserving mechanisms — as the work matures.
The argument here is not that we have built any of this. The argument is that the conceptual frame for what should be built, what should not be built, and where the substrate should sit relative to existing institutions, is now clear enough to publish.
— Rehan Temkar, Co-founder, Apik Systems