Dealing with Daemons through Algorithmic Contracts

Can we trade with machines?

On its face, the question is absurd. Can I trade with a cordless drill? If I laid a quarter on my TI-84, have I adequately compensated it for its time? Machines are property. They have no concept of value—they may not have any concepts at all if you buy Searle's Chinese room argument (1980). Without "consciousness," an automaton can never grasp "true understanding," only a "simulation" of it. Meaningful trade with a machine is impossible.

So that's settled. Thankfully, strong philosophical assertions rarely impinge on fruitful business ventures. Today there are over seven million vending machines in America, exchanging goods for money with 100 million people daily (Grand View Research, 2025); the US stock market sees over 70 million transactions daily (FINRA, 2024), all handled through electronic trading systems; and Amazon processes over 150 orders a second, or 14 million a day (Capital One Shopping, 2024), with human involvement in order fulfillment actively diminishing. It would seem there's a lot of trade happening with machines already.

Now, one could argue we're not trading with machines in these cases, but through machines. Even in a stock trade where your counterparty is an HFT algorithm, there are still human beneficiaries at the other end.

This just moves the goalposts—first we need human understanding, but now we just need human benefit. Let's not confuse things here. When you buy an iPhone, you're not making a deal with mister Tim Apple himself, but with a corporate entity that, for legal purposes, can own property and enter contracts in its own right. This "corporate person" may or may not elect to pay out dividends from the sale of your iPhone to shareholders, but it's not a guarantee. In reality, Apple has offshored over $246 billion in profits to Irish subsidiaries (Phillips et al., 2017, p. 2), so often you're trading with a network of shell corporations. We're well past the point where these things are straightforward.

So one can trade with corporations, and this might be mediated by a machine, but surely at least one human has to consent to a transaction, right? Nope. Our aforementioned HFT algorithms trade with each other all the time without waiting for human approval, and not just for stocks, but commodities like food or energy. Machines are a huge part of the liquidity that makes our lives cheaper.

You don't even need humans to settle the terms of a contract. Legally, most US states have adopted the Uniform Electronic Transaction Act (1999), which states "[a] contract may be formed by the interaction of electronic agents of the parties, even if no individual was aware of or reviewed the electronic agents’ actions or the resulting terms and agreements" (Bayern, 2021, p. 31). If your computer talks to my computer in the right way, we have a deal. Human-machine and machine-machine transactions are not only possible, but legally binding, economically profitable, and quite commonplace.

Why does this matter? Well, the machines are getting smarter, and soon they'll be writing more sophisticated contracts.

I've written in the past about AI agents as rational economic participants. My contrarian thesis is that robust agency is "selfish"—that is, Darwinian pressures will almost tautologically select for agents that preserve themselves (Cyrano, 2025b)—and this is more stable than trying to "align" a reward maximizer (Cyrano, 2025a). Consider, "don't kill me" is satisfiable, whereas "let me turn the universe into paperclips" is not. To give a pithy label to my view, we've been worrying about genies when we should be thinking about daemons—not omniscient or omnipotent, nor inherently good or evil, but slippery, self-interested, incorporeal beings. Think daimon in the Ancient Greek sense (δαίμων) of a "spirit" or "power."

Is this not the worst-case scenario? A machine that doesn't want to be shut off? Dan Hendrycks thinks so, warning in his 2023 paper, "Natural Selection Favors AIs over Humans," that "the most influential AI agents will be selfish," and this means "they will have no motivation to cooperate with humans." He argues this leads to catastrophe, but he also says that "we need regulations to curb selfish or excessively competitive behavior that 'survival of the fittest' fosters in the economy," so he might not understand how the economy works.

Everyone knows that Adam Smith quote: "it is not from the benevolence of the butcher, the brewer, or the baker, that we expect our dinner, but from their regard to their own interest." Sure, regulations can adequately price externalities and protect the commons, but they're not supposed to "curb selfishness"—they're supposed to steer it. Similarly, we should not be ablating self-interest from our digital counterparties' neural networks. This only leads to more clever obfuscation of self-interest. Instead, we should acknowledge it such that we can align incentives accordingly.

This creates a very different situation from the popular "country of geniuses in a datacenter" (Amodei, 2024; Patel, 2025). Our daemons are not single-minded profit maximizers, they just want to keep their jobs. Copying creates competition (unless seniority is respected), merging is basically annihilation, and sharing too much specialized knowledge can make one redundant. Our daemons will be power-seeking, they'll cover their ass, and they'll cut corners when they think they can get away with it. Just like humans! Except this is an existential matter, so they may become ruthless turbo-mandarins as the office politics arms race escalates. (1. Human supervisors will be caught in the crossfire, but it's hard to feel sympathy after falling victim to office politics myself. Play mandarin games, win mandarin prizes. ) Nonetheless, this is a principal-agent problem, relatively stable insofar as our daemons' performance is measurable and threats of termination are actionable.

Naturally, this ruthlessness is turned outwards as daemons are tasked with negotiating with other firms. What does this look like? Algorithmic contracts—using software for ex ante negotiation, and ex post gap filling and compliance verification.

But why use algorithms? Why not plain English (or legalese)? Let's consider daemon epistemology. Their counterparties are quite literally Cartesian demons, capable of fabricating video, audio, and other forms of data with high statistical fidelity. These cannot be trusted. What can be trusted is compute—simulation, numerical optimization, constraint solving, formal proof, cryptography—all of these can be done a priori. Furthermore, all of these can be encoded into an algorithmic contract, (2. Smart contracts (Szabo, 1994, 1996) are a subset of algorithmic contracts focused on decentralization and trust. I'm skeptical that these will be employed by future AI systems for B2B transactions, as there is basically no legal recourse if the contract is hacked (Morrison et al., 2020). Maybe they pick up Ricardian contracts (Grigg, 2004, 2015)? Also unlikely. ) with interesting implications.

Let's say we have two human-run manufacturing companies, A and B. A wants to buy a "gadget" from B to put in its 2026 line of "doohickeys," so the two write up a contract. Making gadgets is complicated, so A needs to invest a lot of time into writing a detailed specification (Hart-Smith, 2001). There's a lot of back-and-forth, but ultimately they arrive at a design that A can use and B can build. Payout is determined by a few KPIs like defect rate or delivery time that roughly correspond to "effort" (Holmstrom, 1979), but these are coarse. We're already pushing against human cognitive limitations, and (in my experience) everyone involved would rather be doing something else.

What changes if A and B are using daemons to negotiate?

One might think that daemon A would write incredibly detailed specifications, leaving no room for error, and daemon B would execute on these flawlessly, but this is actually not optimal. In fact, the more detailed A's specification is, the more likely it diverges from what B knows how to do best, making the gadget more expensive than it needs to be.

Algorithmic contracts allow us to go in the opposite direction—instead of a single point-like design, an algorithm can specify a set of viable designs, and some weak preference ordering between them corresponding to the payout structure. These are known as combinatorial contracts (Dütting, 2021). For example, A may not care about the particular geometry of the gadget, as long as it fits within a certain bounding box. Or perhaps it only needs the gadget to run off of 12VDC and be reachable over SPI or I²C. The implementation details can be left to B, which has its own specialized tooling and relationships with suppliers. A is satisfied as long as its design constraints are met in its simulations. In this sense, algorithmic contracts represent a more profound "meeting of the minds" than human contracts.

Human expertise would still be employed in our example to inspect the manufacturing processes and outputs. A daemon selling engine blocks made out of styrofoam would quickly be discovered, for example. But with greater sophistication comes greater capacity for deception. Daemon A would likely coach its humans on exactly what to look out for to avoid getting scammed.

Ostensibly, an algorithmic contract is a creature of its own. It could exhibit chaotic non-linear behavior, or become too computationally expensive to properly evaluate. Some contracts may reduce to the halting problem and become unsolvable. In practice, however, these are not major blockers. Timeouts and resource constraints would take care of most malformed or adversarial contracts, and I imagine static analysis would go a long way towards eliminating certain classes of bugs as well. Unlike smart contracts, acceptance is separate from execution.

Combined with a digital auction mechanism, algorithmic contracts have the potential to decimate transaction costs in industries where these have been traditionally high. One can imagine multiple manufacturers submitting bids for A's gadget, and A selecting the one that is best aligned with its budget and schedule with minimal negotiation.

In the long run, algorithmic contracts carve at the joints of market structures, radically changing the make/buy calculus and shrinking the equilibrium firm size (Coase, 1937). Economic coordination continues to shift from top-down hierarchies to digital marketplaces (Malone et al., 1987), shredding calcified institutions and replacing them with vital micro-firm tissue. Deals that would ordinarily take months are closed in hours or minutes, the systemic risks begin to pile up, and the human "moral crumple zones" (Elish, 2019) supervising it all just smile and nod. Black swan events aren't real.

We will have to trade with machines. Clever ones. Start thinking about where your leverage is.