The autonomous economy has spent enormous energy on one question: how do we keep agents from doing the wrong thing?
Permissions. Spending limits. Sandboxes. Monitoring. Human approval. Kill switches.
All of these matter. None of them is enough.
Even a well-designed agent will eventually create a dispute. It will misunderstand a mandate. Act on stale authority. Buy from another malfunctioning agent. Pay for something that never arrives. Exceed a budget by $3,000. Delegate a task it was never authorized to delegate. Accept defective performance. Make a reasonable decision that its principal later contests.
At that point, the problem is one of law.
Who had authority? What could the counterparty reasonably rely on? Which rules governed the transaction? What evidence controls? Who decides? Once a decision is made, who gets the money?
The agent economy needs institutions around its agents. It needs private law for machines.
This does not mean giving software citizenship or putting robots on trial. It means using private law in its older, practical sense: participants choose the rules for their relationships, choose a forum for disputes, and give that forum bounded power to grant remedies when things go wrong.
Authority has to become legible
Software already acts for people and businesses without a complete account of its authority. An agent may hold a wallet, API key, corporate account, procurement login, or payment credential. Possessing a credential does not answer what the agent may do with it.
An agent should enter commerce with a clear mandate: who authorized it, what it may do, for how long, subject to what limits, and whether it may delegate.
The power to recommend a purchase is different from the power to make one. The power to spend $500 is different from the power to spend $50,000. The power to order goods may not include the power to accept defective performance, waive a claim, or settle a dispute.
Those limits need to follow the task across agents and services. At minimum, the transaction needs records that show what authority existed when the agent acted. Cryptography can prove that a key signed a message. It cannot prove that the key was entitled to do what it did.
Commerce needs a failure path
Most agentic-commerce systems are designed around the happy path. An agent receives a task, finds a service, exchanges credentials, pays, and receives performance.
Mature commercial systems are defined by what happens when that path breaks.
Credit cards have chargebacks. Marketplaces have seller disputes. Banks have error-resolution procedures. Contracts have courts and arbitration clauses. Autonomous commerce needs its own failure path.
Before a transaction, participants should be able to adopt common rules and choose a neutral forum. Those rules should say which disputes the forum may hear, what evidence the parties must preserve, how each side may challenge a claim, and which remedies the forum may issue.
Start with a narrow jurisdiction. A private rulebook binds the participants who adopt it and the transactions they place within it. It does not bind an infrastructure provider that never joined or displace mandatory public law.
A transaction should carry its legal context along with its payment instruction:
These are the parties. This is the agent’s mandate. These are the governing rules. This is the evidence record. This is where a dispute goes. These are the available remedies.
A kill switch is not a court
Agent-safety systems need emergency controls. Revoke the credential. Freeze the account. Stop the payment. Disable the capability. Shut the agent down.
Sometimes that is exactly the right response. It still does not decide the dispute.
A kill switch can prevent the next transaction. It cannot decide whether the last transaction was authorized, whether a merchant reasonably relied on the authority presented, whether promised performance occurred, or whether money already paid should be returned.
Those are backward-looking questions. They require evidence, notice, a chance to respond, a rule of decision, and a remedy. They require adjudication.
Platforms, payment systems, and identity providers may need emergency powers. They should not each have to become a court. A separate dispute layer lets infrastructure providers remain infrastructure providers.
Recourse is part of safety
Commercial systems do not become safe by assuming that contracts will never be breached, payments will never be fraudulent, or employees will never exceed their authority. They become safe by absorbing failure.
For agents, that means alignment before the act and recourse after it.
Recourse may mean a refund, release of escrow, correction of a transaction record, restoration or suspension of a capability, reassignment of funds, or compensation from the participant that accepted the relevant obligation.
The basic constraints are familiar. The claimant proves a recognized claim. The other side responds. The decision explains the result. The remedy reaches only parties, assets, and systems within the forum’s authority.
Dispute resolution is part of the safety architecture for agentic commerce.
Judgment can become infrastructure
Structured decisions can improve the system that produced the dispute.
Repeated cases may show that a mandate standard is ambiguous, a verification service is unreliable, a delegation rule creates unexpected risk, or merchants interpret a field differently from agent developers.
Reasoned decisions make those problems visible. They show builders which language causes disputes, tell counterparties which representations they may rely on, and show networks where losses are likely to fall.
The forum decides each dispute under the rules the parties adopted. The participants change those rules prospectively. Over time, the resulting body of decisions turns individual failures into clearer expectations and better system design.
Start with a small jurisdiction
Start with a defined group of businesses, a defined class of transactions, known participants, shared rules, preserved evidence, a neutral forum, and bounded remedies.
That is enough to create a small legal system.
People’s Court is building an AI-native private dispute-resolution layer for commerce. It can receive transaction records and evidence, hear both sides, apply agreed rules, issue reasoned decisions, and turn those decisions into practical remedies where the surrounding infrastructure permits.
As autonomous agents begin to transact with merchants, platforms, payment networks, and one another, the central governance question changes. We still need to ask how to control what an agent may do. We also need to ask who decides what happens after it has already acted.
Every serious commercial system develops an answer. The agent economy will too.