Trust between strangers is not natural. We forget this because the machinery that makes it work has become invisible. When you hand a card to a merchant you have never met, you are not trusting the merchant. You are trusting a bank, a card network, a legal system, and a chargeback process that will claw your money back if the goods never arrive. Strip that machinery away and you would not hand a stranger anything.
For most of history, trade with strangers required an institution in the middle. Merchants invented letters of credit so a buyer in one city could pay a seller in another without either trusting the other. Courts made contracts worth signing. Platforms built reputation systems so a buyer could size up a seller from a star rating. Every one of these is the same move: outsource the trust to something bigger than the two parties.
Now put a software agent on each end of the transaction and watch every piece of that machinery fail.
The Machinery Assumes a Person
An agent cannot sign a contract in any sense a court recognizes. It has no legal personality, no assets, and no fear of consequences. You cannot sue a process. You cannot repossess anything from a piece of code that finished executing an hour ago.
An agent also cannot size up a counterparty the way people do. My years in enterprise security taught me how much of human trust is pattern recognition built over a lifetime: the tone of an email, the age of a domain, the feel of a negotiation that is slightly off. Agents have none of this context, and worse, everything they might use as a signal can be fabricated at zero cost. A convincing website, a warm reply, a plausible track record. Machines can generate all of it faster than machines can evaluate it.
Reputation systems fail too. Reputation works when identity is scarce and history is expensive to build. An agent can be instantiated a thousand times before lunch. A five star history means nothing when the seller can discard the identity the moment it becomes a liability.
So What Is Left
Here is the uncomfortable conclusion I keep arriving at: when both parties are code, trust in the human sense is simply not available. There is no belief, no benefit of the doubt, no relationship that thickens over time. Asking whether one agent trusts another is like asking whether a compiler trusts the source code. The question is malformed.
What is available is something narrower and, I would argue, better. An agent can verify. It can check a cryptographic signature. It can confirm that funds are actually locked before it does the work. It can evaluate whether a delivered result matches an agreed specification. None of this requires believing anything about the counterparty's intentions, because intentions stop mattering when the mechanism makes defection unprofitable or impossible.
This is the same shift security went through. We spent decades trying to decide which networks and devices to trust, and the answer that finally worked was to stop deciding. Assume nothing, verify everything, and let access follow from proof rather than position. The industry called it zero trust. It was really a confession that trust never belonged in the architecture in the first place.
Designing for the Absence of Trust
If agents are going to transact with strangers, and they will, the environment they transact in has to supply what institutions supplied for humans: a way to make commitments binding without anyone believing anyone. Agreements that execute rather than persuade. Payments that settle when conditions are met rather than when someone decides to honor an invoice. Identity that is proven at the moment of interaction rather than assumed from history.
This is the problem I chose to work on when I started Setix, and it is why I think of the work as infrastructure rather than product. The interesting question is not how to make agents trustworthy. It is how to build an economy where trustworthiness is irrelevant, because every claim that matters can be checked by the party relying on it.
Machines do not do trust. Once you accept that, you stop trying to teach them something they cannot learn and start building the thing they actually need: a world where verification is cheap, universal, and final. Humans might come to appreciate living in that world too.