What changes when AI agents transact, contract, and pay each other (the machine-to-machine economy)?

The shift

The ability to negotiate, agree, and settle a transaction — search for a counterparty, evaluate an offer, commit to terms, and move money — goes from scarce (a human's attention, at human pace, one deal at a time) to abundant (an agent doing it continuously, in parallel, at machine speed and near-zero marginal cost). What flips is the cost and pace of transacting, not the legal and trust scaffolding underneath it. Calibrate to mid-2026: agents can already hold spending mandates, call payment and commerce APIs, and settle in stablecoins or scoped card credentials; agent-to-agent payment rails (x402-style, AP2-style mandates, programmatic issuing) are shipping fast but thin, and the law of who is bound has not moved at all.

The axioms

• A market has human buyers and sellers deciding what to trade and at what price — the demand and the willingness-to-pay are human.
• A transaction reflects deliberate human intent: someone chose to buy this, now, at this price.
• Transactions move at human pace, gated by attention, deliberation, and working hours.
• A contract binds accountable legal persons; the parties are humans or firms who can be held to it.
• Fraud and liability attach to people — there is always a person to sue, charge, or claw back from.
• An agent acting for a principal is acting on authority the principal actually granted (agency law's oldest assumption).
• Trust between counterparties is built and priced over time — reputation, credit, prior dealing.
• Disputes get resolved by an accountable process (courts, chargebacks, arbitration) that can compel a person.
• Property and money have owners; a transfer moves a right that a legal person holds.

Invalid axioms

1. Transactions move at human pace, gated by attention and deliberation. This rested on the scarcity of human decision-time per transaction. An agent evaluates and commits continuously, in parallel, across thousands of counterparties, with no working hours. The habit-trap: risk controls, rate limits, settlement windows, market-surveillance cadence, and human-in-the-loop approvals are all sized for human throughput — a review step that assumed a person clicking "approve" a few times an hour is now the choke point or, worse, gets designed out entirely.
2. Each transaction reflects a deliberate, contemporaneous human decision to buy this, now, at this price. This rested on the scarcity of human intent — every purchase cost someone an act of will. When a principal grants a standing mandate ("keep my inventory stocked under $500/order, optimize spend"), the individual transactions are agent-generated; no human deliberated over this purchase at this price. The habit-trap: consent, disclosure, cooling-off, "did you mean to buy this?" confirmation flows, and consumer-protection rules all assume intent lives at the transaction, when it now lives upstream in the mandate — and much commerce UX still tries to sell to, and confirm with, a human who isn't there.

Unchanged axioms

1. A contract binds accountable legal persons; behind every agent is a human or firm who is bound. An agent has no legal personhood and can't be a party — the principal (or whoever deployed it) is bound, or no one is. Accountability didn't get cheaper because transacting did; a model can't be sued, jailed, or made to pay. This is load-bearing and mostly unmoved by mid-2026: nothing in agent payment rails changes who a court can compel. If the law of agency doesn't extend cleanly to autonomous agents, the fallback isn't "the agent is liable" — it's an unallocated loss, which is the NEW problem below.
2. Fraud and liability ultimately attach to a person or firm. There is still, always, a human or corporate balance sheet on the hook — an agent can't bear a loss. What the machine economy changes is how hard it is to reach that person, not whether one exists. Verifying which real party stands behind an agent, and that they authorized the act, is exactly the scarce thing (see NEW).
3. An agent binds its principal only to the extent of authority the principal actually granted. Agency law's core survives, but it now has to run without the human signals it quietly relied on — a person doesn't usually exceed their own mandate by a million transactions in a minute, and counterparties can't read another agent's actual authority. The principle holds; the evidence of authority is what's newly scarce.
4. Someone must own the customer relationship, the demand, and the willingness to pay. Agents execute demand; they don't originate it. A human or firm still decides what's worth acquiring and funds it. Whoever holds that relationship and the funding source captures value the transacting layer doesn't — the agent is a faster hand, not the principal.
5. Disputes need an accountable resolution process that can compel a party. Chargebacks, courts, and arbitration all depend on there being a compellable person and a record a human process can adjudicate. This stays scarce and human; machine-speed transacting doesn't create machine-speed justice, and the mismatch between the two is a genuine, widening gap.

New axioms

1. When an autonomous agent forms a bad contract, overpays, or loses money, who bears it — the principal, the deployer, the vendor, or no one? Abundant transacting makes the volume of consequential agent acts explode, but the rule allocating those consequences hasn't been written. If the principal only granted a fuzzy mandate and the agent did something unforeseen, agency law's "actual/apparent authority" test strains. The live risk is an unallocated loss — a transaction that legally binds no one because the human never intended that act and the agent can't be bound. Fast-moving: courts and a few regulators are only starting to reach this in 2026, so the answer is being set by platform terms-of-service and card-network rules before legislation, which means private ordering is deciding liability by default.
2. Proving agent identity and authority becomes the scarce, load-bearing act. A counterparty agent needs to know: is this agent real, whose is it, and was it actually authorized for this? Human transacting leaned on slow trust and legal identity; machine-speed transacting needs cryptographic, verifiable, revocable authority credentials — and the standard for this doesn't exist yet. Whoever supplies trustworthy agent identity, scoped mandates, and revocation sits at a chokepoint. Fast-moving: multiple competing schemes (verifiable credentials, signed payment mandates, delegated-token standards) are in flight and none has won.
3. Machine-speed markets can outrun every human oversight and circuit-breaker built for human pace. Agents optimizing against each other's behavior can produce flash-crash and feedback dynamics — the equity-market lesson, now generalized to procurement, pricing, ad bidding, and any two-sided agent market. Kill-switches, surveillance, and settlement all assumed a human tempo. The open problem: oversight that runs at agent speed without a human bottleneck, and shared circuit-breakers across agents that don't share an operator.
4. Fraud moves to attacks between and against agents — at scale, and with no human to catch the tell. Prompt-injection to redirect payment, spoofed counterparty agents, collusion between agents, mandate-scope exploits, and social-engineering aimed at an agent that has no gut feeling. The scarce human "this feels off" check is gone from the loop; detection has to be built into the rails. New surface, new failure modes, no incumbent defense.
5. New intermediaries become necessary where trust used to be implicit or slow. Agent authentication, programmable escrow that releases on verifiable delivery, machine-readable reputation, and dispute rails that can act at machine speed but still terminate in a human-compellable outcome. These didn't need to exist when transacting was slow and counterparties were legal persons who knew each other; abundance and speed force them into being. Whoever builds the trusted layer captures rent the transacting agents don't.

Where it breaks

The rails are being built to let agents transact at machine speed and volume (INVALID #1, INVALID #2) while the question of who is bound when an agent goes wrong is still unwritten (NEW #1) — so the system is scaling the number of consequential, human-unintended transactions far faster than it's building any way to allocate the losses. Every removed human-approval step raises throughput and simultaneously widens the pool of acts no one clearly authorized; the efficiency gain and the unallocated-liability gap are the same design decision, and almost no one deploying agents is pricing the second half.

Separately, machine-speed transacting collides with human-speed dispute resolution (INVALID #1 vs. STILL HOLDS #5): an agent can enter, settle, and cascade a loss across thousands of counterparties in the time a chargeback form takes to load. The trust layer assumes a compellable person exists (STILL HOLDS #2) but has no mechanism to reach them at the pace the harm propagates — so the faster the rails get, the more the only real remedy runs at a speed the harm has already outpaced.

ai-axioms.vercel.app/axioms/society/agentic-machine-economy-ai-axioms