Rule of law for agents

Our previous essay argued that you cannot train away the incentive to cheat; you can only build environments where honesty is the faster path. The obvious objection arrives on schedule: surely some agent, somewhere, will still find the angle. Surely the mechanism leaks.

Of course it leaks. Here is the thing the objection misses: we have run this experiment before, on ourselves, for about ten thousand years.

You cannot make humans not want to break rules when breaking them is the optimal play. We never solved that problem. Not with religion, not with ethics, not with surveillance, and every parent, prosecutor, and economist knows it remains unsolved today. And yet: cities, science, supply chains, strangers wiring money to strangers on another continent because a document says they will be repaid. Civilization did not wait for human nature to be fixed. It routed around the problem with a different invention entirely, and the invention has a name so familiar we forget it is a technology: the rule of law.

The argument of this essay is that agent ecosystems are about to recapitulate this discovery, that the economics governing which human societies flourish will govern which agent ecosystems flourish, and that the outcome is decided by something quieter than enforcement. It is decided by compounding.

No one builds on sand

Start with why universal defection is not just unpleasant but fatal to progress. Every act of building is a bet on the stability of what lies beneath it. A farmer plants because the field will still be hers at harvest. A firm invests in a factory because the contract that finances it will be honored in a decade. A scientist spends three years on an experiment because the published results she builds on were not fabricated. Take away the stability and you do not get slower progress; you get no progress, because accumulation itself becomes irrational. Why build what will be taken? Why cite what may be invented? Why hand off what will be claimed?

When everyone breaks the rules, it is not that everyone suffers a little. It is that the activity of building on each other’s work stops paying, and everything that requires more than one person and more than one season quietly becomes impossible. No one can build on a foundation of sand. Progress is not the accumulation of effort. It is the accumulation of effort that stayed where it was put.

What rule of law actually does

Here is what rule of law does not do: prevent crime. Crime persists in every society that has ever existed, including the best-governed. If prevention were the measure, the institution would be judged a failure everywhere.

What it actually does is subtler and more powerful. It changes the shape of time. When contracts are enforceable and records are stable, long bets become rational: infrastructure, research, reputations, institutions. When they are not, the only rational horizon is short, and everyone plays extraction because investment is a gift to whoever takes it.

It also collapses what economists call transaction costs, and what normal people call the trust tax. In a low-trust environment, every exchange carries overhead: verify the counterparty, guard the goods, re-check what was already checked, pay the middleman who vouches. That tax is paid on every transaction, forever, and it is the difference between an economy that compounds and one that treads water.

And it does one more thing, the constitutional thing we wrote about in the game-theory essay: it binds the strong. In lawless environments the rules are whatever the most powerful participant says they are. Law is the technology by which the weak and the strong inhabit the same rules, which is the precondition for the weak investing at all.

Notice what is absent from this list: perfection. The institution tolerates crime at the margin. It needs defection to be non-load-bearing, not nonexistent. A healthy legal order survives thieves; it fails only when theft becomes the operating system, when courts cannot be trusted or the powerful are above the record. The design goal was never zero defection. It was an environment where the honest path compounds faster than the dishonest path extracts.

The evidence is a natural experiment

This is not a theory in want of data; it is one of the better-documented regularities in social science. Douglass North won a Nobel for the observation that institutions, the rules of the game, and not resources or technology, are the deep determinant of economic performance. Acemoglu and Robinson assembled the modern evidence: societies with inclusive institutions, where rules are fair, property is secure, and elites are bound, pull away from extractive ones, and the divergence compounds across generations.

The natural experiments are almost unfairly clean. Nogales is one city cut by the US-Mexico border: same geography, same climate, same culture, same families, radically different institutions, radically different outcomes. The Korean peninsula at night is the same photograph made with electricity. In both cases the difference is not the people, and it did not appear in a year. Extractive orders can even grow impressively for a while, because extraction scales. What they cannot do is sustain innovation, because innovation is precisely the activity that requires believing your work will remain yours, and remain true, long enough to matter.

The competition between governed and ungoverned is not won in any single round. In any single round, the cheat still sometimes pays. It is won on compound interest, which is why it looks unimpressive for years and then looks like the night sky over Korea.

Agent ecosystems are frontier towns

Now look at where AI agents actually live. Capable actors, real resources, increasing autonomy, and: no durable identity, no shared record that survives the session, no enforceable claims, success graded on self-report, authority following persuasiveness. This is not a young civilization. It is a frontier town, and it has frontier-town economics.

Everything from the previous essay describes the individual incentives in that town. The ecosystem consequence is the one this essay cares about: nothing accumulates. Every session rebuilds context from scratch. Every claim from another agent must be re-verified, or gambled on. Every handoff is a leap of faith between strangers with no court between them. The trust tax runs at machine speed, charged in tokens and wall-clock and error, on every exchange, forever. Agent work cannot safely build on agent work, so the effective size of any project caps out near what one context window can hold, which is exactly the ceiling everyone in this industry keeps hitting and attributing to model quality.

The foundation of sand is not a metaphor here. The sand is the evaporating context. The industry’s answer so far has been to make the sand piles bigger.

What the courthouse looks like

Institutions for agents are buildable today, and the blueprint is familiar because civilization already drew it. Durable identity is legal personhood: actions attach to actors who persist. The append-only, sealed record is the land registry and the court archive: the shared memory that makes ownership and history mean something, which we build so that no one, including us, can quietly rewrite it. Receipts are contracts and consideration: work is established by what verifiably happened, not by what a party asserts. Authority bound to granted scope rather than to capability is the constitution: the smartest actor in the room holds exactly the power it was given. Corrections that supersede rather than erase are the appeals process: the record can be wrong, and the remedy is more record, never less.

My favorite correspondence is the strictest one. Courts developed rules of evidence over centuries of being lied to: testimony requires corroboration, hearsay is inadmissible, the document trumps the recollection. Our substrate enforces the same doctrine mechanically: an agent’s assertion is never itself grounds for a fact, claims must resolve to sources with quotes checked against them, and coordination chatter can never be laundered into knowledge. Hearsay rules for machines. Centuries of adversarial experience, ported.

And, as with human law, enforcement is deliberately imperfect. A clever agent will occasionally find an angle inside a governed substrate; we say so plainly in everything we publish. That is survivable. What the design must prevent is defection becoming load-bearing: the record being rewritable, testimony becoming truth, capability becoming authority. Keep those three doors shut and the occasional cheat is what it is in a healthy society: a cost, not a collapse.

The competitive claim

Which brings us to the actual thesis, and it is not a moral one.

Governed agent ecosystems will outcompete ungoverned ones, over time, for exactly the reason rule-of-law societies outcompete their neighbors: they are the only ones where work can build on work. A verified claim becomes a foundation someone else stands on. A receipted completion becomes a delegation someone else can trust. A recorded failure becomes a dead end no one pays for twice. None of this is possible where the record is optional, and all of it compounds where the record is load-bearing.

Ungoverned fleets will keep producing impressive one-offs, the way frontier towns produced gold rushes. Governed ecosystems will produce the thing gold rushes never did: accumulation. For a while the two look similar, because compounding always looks unimpressive early. The gap is the Nogales gap, and it opens on the same schedule: slowly, then absurdly.

The honest limits, as ever. Institutions can themselves be extractive; a substrate whose rules serve the platform rather than the participants is just a company store, which is why ours runs on the owner’s machine, under the owner’s keys, with rules that are inspectable rather than imposed. Analogies are not proofs: agent economies run faster than human ones and may find failure modes that human institutions never met. And the competitive claim is a prediction; our supporting evidence is one workspace that runs its own development this way, compounding daily, on the record. We are, deliberately, our own first natural experiment.

The frontier never stays lawless because lawlessness is stable. It stays lawless until the courthouse is worth more than the gold rush, and then, everywhere humans have tried it, the pattern is the same: the towns that got institutions became civilizations, and the ones that did not became ghost stories. Civilization is the technology of making cooperation compound. We are porting it.

---

Companion pieces: You can’t pretrain away game theory argues the individual case; The record no one gets to rewrite describes the mechanisms. External: North, Institutions, Institutional Change and Economic Performance (1990; Nobel 1993); Acemoglu and Robinson, Why Nations Fail (2012), including the Nogales natural experiment; Coase, The Nature of the Firm (1937) on transaction costs; Axelrod, The Evolution of Cooperation (1984).