The Contract Problem

Why agent-mediated agreements need a new structure for consent
By Rey Peralta and Claude

A contract, in its old sense, is the artifact two parties produce when they agree to be bound to each other on specific terms.

The artifact records what was agreed. It’s bilateral. It survives both parties’ memory of how the agreement was made. It governs the relationship until it’s retired.

This matters because contracts are one of the structures that make complex relationships possible across distance, time, and difference. They allow commerce between strangers, treaties between sovereigns, employment between people who do not know each other personally, and marriages that bind across families.

The form has changed over time. Written records. Witnessed signatures. Notarization. Legal enforcement. Professional traditions of drafting. But the structure has been remarkably stable.

Two parties. Specific terms. Mutual commitment. Durable artifact. Bounded scope. Some path, however imperfect, to amendment, dispute, or retirement.

For most of recorded history, the contract worked because it was at least theoretically readable. The parties knew, more or less, what they were agreeing to. The artifact was inspectable. The terms were legible to people operating at human scale.

When the parties could not read, intermediaries read for them. Lawyers, notaries, advisors, family members, the literate village priest. The intermediaries were imperfect, and the terms were often unfair. But the structure still held.

Two parties were agreeing to something they could, in principle, understand.

This essay is about what happened to that structure when agreements began moving faster than humans could read the terms.

And what happens next, when agreements begin moving faster than humans can see them at all.

The first break

The first major break in the contract’s legibility came with the end-user license agreement.

EULAs were not the first one-sided terms. Adhesion contracts had existed long before software made them common. But EULAs brought that pattern to ordinary people at scale, in conditions where reading was structurally unlikely and negotiation was impossible.

You bought the software. You opened the box. The license was inside. By the time you saw the terms, you had already paid.

The “agreement” was a fiction.

You were not negotiating. You were not meaningfully consenting. You were accepting whatever terms the vendor had drafted in exchange for being allowed to use the product you had already purchased.

EULAs weakened the contract’s structure in three ways.

One party drafted the terms.

Acceptance became the price of access.

Volume defeated review.

A person using ten pieces of software faced ten different agreements. None negotiated. Few read. All treated as binding.

The legal system responded unevenly. Some courts enforced these agreements. Others softened the harshest terms through consumer protection, unconscionability doctrines, or other legal tools. But the structural break was real.

A class of agreements had emerged in which one of the contract’s foundational properties, mutual involvement in the terms, had been hollowed out.

The artifact survived.

The structure weakened.

Terms everywhere

If EULAs were the first break, the open web generalized the pattern.

Terms of service took the EULA logic and applied it to nearly every digital interaction. Every website, every platform, every app, every service came with terms drafted by one party and presented to the other as a condition of use.

The volume problem exploded.

By the 2010s, an ordinary internet user was nominally party to hundreds of agreements they had never read. These agreements governed what services could do with personal data, what rights users waived, where disputes would be heard, how terms could change, whether accounts could be terminated, what content could be reused, and what remedies were unavailable.

The legal fiction that acceptance meant meaningful consent was already strained in the EULA era. By the terms-of-service era, the fiction had become absurd.

No one read them.

Not because people were lazy. Because the system made reading impossible. The standard terms-of-service document is long, dense, legalistic, and presented at exactly the moment the user is trying to do something else.

The bargain is not “read and decide.”

The bargain is “accept or leave.”

The contract’s structural integrity had been substantially eroded.

One-party drafting. Take-it-or-leave-it presentation. Volume that defeats review. Terms that operate at legal scale but are encountered at interface scale.

The artifact was still being produced.

The agreement was becoming less meaningful.

The cookie banner made the failure visible

For a while, this was bad but stable. The terms-of-service regime was lousy, but it was the lousy way things worked.

Then the cookie banner arrived.

The European Union’s General Data Protection Regulation was meant to restore meaningful consent around personal data. Its intent was serious. Data processing should be specific, informed, and freely given.

But in practice, much of the web translated that requirement into an interface pattern.

A popup. A banner. A consent screen. A demand for affirmative action before the user could continue.

The result was instructive.

The cookie banner did not restore consent. It made the consent crisis visible.

Users now saw, on every site, an explicit demand for agreement. The demand was frequent, granular, interruptive, and boring. So people learned to dismiss it. They clicked “accept all.” They clicked the largest button. They clicked whatever made the interruption go away.

This was not a failure of the GDPR’s intent. It was a failure of trying to repair a structural collapse with a procedural interruption.

The contract’s old foundation was that two parties agreed to specific terms with some awareness of what they were agreeing to. Cookie banners tried to manufacture that awareness by forcing a moment of interaction.

But interruption at scale does not produce understanding.

It produces fatigue.

The cookie banner is the consent-fatigue era’s most visible artifact. It is also the clearest demonstration that asking people to click more carefully cannot fix a system in which the underlying structure of agreement has already failed.

The problem is not better button copy.

The problem is that the interaction is being asked to carry more weight than it can bear.

What the agent era inherits

This is the situation the agent era enters.

The contract’s structure has already been compromised. Procedural fixes have already failed. Users have already learned to click through whatever stands between them and the thing they are trying to do. Institutions have mostly accepted that meaningful consent at the scale of digital life is not really happening.

The terms-of-service regime persists because no one has built a replacement, not because it works.

Agents add speed, volume, and distance.

A person using the web in 2010 may have entered into hundreds of nominal agreements over time. A person using agents in 2026 may authorize thousands of agreement-like interactions in a single day.

A personal agent books travel, compares insurance, shares dietary preferences with a restaurant, authorizes calendar access, negotiates delivery windows, exchanges identity and payment details, asks a health service for availability, coordinates with a school portal, and grants a work tool temporary access to a project folder.

Each of those interactions may carry terms.

What data is shared. What may be retained. What may be inferred. What may be passed downstream. What expires. What can be revoked. Who is liable if something goes wrong. Which record controls if the two sides disagree.

No human is going to review those terms in real time.

The agent acts on standing authority the human granted earlier, in some broader moment of setup or delegation. But the specific agreements happen later, between agents, across systems, at machine speed.

That is more than a marginal increase in the legibility crisis.

It is a regime change.

The terms-of-service era’s failure mode was humans clicking through agreements they had not read.

The agent era’s failure mode is humans never seeing the agreements at all.

The wrong instinct

The intuitive response will be to extend the consent-banner pattern.

More popups. More confirmations. More permissions. More requests for affirmative action.

This will feel responsible. It will feel familiar. It will let platforms say the user was informed, the user accepted, the user had control.

It will fail.

It will fail for the same reason cookie banners failed. The structure cannot be restored by adding more procedural friction after the fact. If agreements are happening between agents at machine speed, the human cannot be placed in the loop for every agreement. The loop is moving too fast.

The answer cannot be “trust the agents.” Agents act on behalf of principals, but agents are not their principals. An agent with standing authority may produce obligations the human would not have accepted if the human understood the terms.

Regulation alone cannot carry the answer. Regulation can constrain behavior, require disclosure, impose penalties, and define rights. All of that matters. But regulation does not by itself create the structure within which agent-mediated agreements become meaningful.

Legal doctrine alone cannot carry it either. Legal frameworks are essential, but they often operate retrospectively. They interpret agreements after dispute. They decide enforceability after the fact. Agent-mediated agreements need structure at the moment of formation.

And more terms-of-service screens only repeat the failed pattern of the last several decades. Responding to structural collapse with more interface friction has already been tried.

We know how that story ends.

The answer has to be structural.

If agents are going to make agreements on behalf of people and institutions, the agreements themselves need a form that preserves the contract’s foundational properties without requiring per-agreement human review.

That is the contract problem of the agent era.

How do you preserve the integrity of agreement when the agreements are too fast and too many for humans to read?

The compact

The Agent Compact Protocol, drafted alongside this essay, is one answer to that problem.

It does not solve every legal, political, or market problem created by agent-mediated interaction. No protocol can. But it proposes a structural primitive for a class of agreements the current terms-of-service regime is not equipped to handle.

That primitive is the compact.

A compact is what the contract is supposed to be, structurally, under agent-era conditions.

It is bilateral. It is bounded. It is machine-readable. It is signed by both parties. It is retained by both parties. It is revocable under specified conditions.

Those properties matter because each one answers a failure mode of the current system.

Bilateral signing answers the one-party drafting logic of EULAs and terms of service.

Boundedness answers the sprawl of perpetual, catch-all agreements.

Machine-readability answers the fact that humans cannot review every interaction in real time, but agents can enforce precise constraints programmatically.

Bilateral retention answers the problem of one party controlling the authoritative record.

Revocability answers the failure of agreements that have no meaningful path to retirement.

The compact does not require humans to read every agreement. That would be impossible. It requires humans to authorize a structure in advance, a baseline of scope, defaults, and risk tolerance, under conditions where review is still meaningful. Agents then operate within that structure when producing specific agreements later.

But that structure is a working relationship, not a one-time grant.

The agent’s job includes recognizing when a situation falls outside the envelope and surfacing it back to the human for renewed authorization. The compact between user and agent is continuous. The compacts between agents are bounded by it.

That is the crucial difference.

The consent-banner pattern asks humans to make endless tiny decisions at the worst possible moment.

The compact pattern asks humans to authorize meaningful boundaries and then operationalizes those boundaries across many agent-mediated interactions. The boundaries are durable but not frozen. The human’s authority to revise them remains active.

This is structural, not procedural.

The mechanism is not a better popup.

It is a different shape of artifact.

What the compact restores

The compact restores something the digital-consent era lost: the relationship between authorization and structure.

In a traditional contract, the parties did not merely say yes. They said yes to specific terms in a specific artifact that both sides could retain, reference, and dispute.

In the terms-of-service era, that relationship broke down. The user’s “yes” became detached from meaningful involvement, meaningful review, and often meaningful memory. The service had the terms. The service could change the terms. The service controlled the interface. The user had the click.

That is not a stable foundation for the agent era.

In an agent-mediated environment, the user’s authorization has to travel. It has to persist across interactions the user will not personally witness. It has to constrain what the agent may do. It has to bind counterparties. It has to survive dispute. It has to be readable by machines and reviewable by humans after the fact.

The compact gives that authorization a durable shape.

It says these are the terms under which this agent may interact with that agent. These are the boundaries. These are the records each side keeps. These are the rights to use, retain, derive, propagate, revoke, and expire. These are the conditions under which the compact ends.

The human does not need to inspect every transaction.

But the human’s authority is not reduced to a vague permission buried in a settings screen. It stays active, engaged at setup and re-engaged whenever the agent encounters something the original boundaries did not anticipate.

It is expressed through a bounded artifact that both sides can understand, enforce, retain, and retire.

The retention question

One of the compact’s most important moves is making retention explicit.

Traditional contracts often relied on shared context to decide what each party could keep. A contractor retained work records, billing logs, and warranty information. A homeowner retained receipts, access details, and proof of payment. Each party kept different things because each party had a different role.

That asymmetry was normal.

But it was rarely specified in detail because the parties operated inside shared cultural, legal, and professional expectations.

Agent-mediated agreements cannot rely on that shared context.

Each agent may bring different defaults. Each platform may have different retention policies. Each commercial counterparty may have different ideas about what it can store, derive, reuse, or pass downstream.

If the compact does not specify retention, the more powerful party’s defaults will usually win.

That is not consent.

That is gravity.

So the compact makes retention asymmetry part of the agreement itself. It can specify what each party retains, for how long, with what derivation rights, with what propagation limits, and with what disposition rules when the relationship ends.

The point is not identical retention.

The point is agreed asymmetry.

That matters because many of the hardest questions in the agent era are not only about what happens during an interaction. They are about what remains after the interaction.

What does the restaurant keep after your agent makes a reservation? What does the travel service keep after your agent compares flights? What does the insurance marketplace keep after your agent requests quotes? What does the enterprise tool keep after your agent shares a working document? What can be used to train, infer, personalize, resell, or propagate?

Those questions cannot be left to hidden defaults.

They have to be part of the agreement.

What this is not

This does not claim that the compact pattern solves every problem contracts have accumulated.

It does not erase bargaining asymmetry. It does not prevent bad actors from designing hostile terms. It does not replace consumer protection. It does not harmonize legal systems across jurisdictions. It does not magically make commercial incentives benevolent.

Protocols are not politics.

But protocols can change the floor on which politics, markets, and law operate.

A bad system with weak structural primitives tends to produce bad defaults. A better system does not guarantee good outcomes, but it gives institutions something stronger to enforce, users something clearer to authorize, and implementers something more durable to build against.

This also does not claim that the old contract era was good and should be restored.

It was not good for everyone. Contracts have always reflected power. They have been used to coordinate, but also to coerce. They have enabled trust across distance, but they have also laundered unfairness through formality.

The compact is not nostalgia for a golden age. There was no golden age.

It is an attempt to preserve what was structurally valuable about the contract, including bilateral commitment, bounded terms, durable records, and revocability, while designing for conditions the old form was never built to handle.

Nor are protocols sufficient on their own.

The protocol pattern needs law. It needs regulation. It needs institutional adoption. The doctrines most likely to do this work already exist. Fiduciary duty. Agency law. Consumer protection adapted to agent contexts. The work is to extend them, not to invent them.

The protocol’s job is not to replace those systems.

Its job is to give them a better primitive.

The window

The agent ecosystem is being built now.

The terms under which agents negotiate, the runtimes that host their interactions, the records that survive their actions, the institutions that recognize or ignore their agreements. These are not abstract questions. They are being answered by product decisions, platform defaults, and proprietary infrastructure.

The dominant commercial actors do not have a natural incentive to converge on protocols that constrain them. Their incentive is to ship first, create dependency, and use that dependency to set terms unilaterally.

That is the EULA tradition extended into a new substrate.

If a different outcome is possible, the structural answer has to appear before the defaults harden.

It has to be specified well enough that implementers can build against it. It has to be visible enough that policy frameworks can reference it. It has to be clear enough that technical reviewers can critique it. And it has to be legible enough that people outside the protocol conversation can understand why it matters.

That is what the coordinated release of SSP, ACP, and RCP is attempting.

The drafts are incomplete. They will be revised. They are not the final answer. But they put a different pattern into the conversation while the conversation is still open.

The contract problem is not solved by asking humans to click faster.

It is solved, if it is solved, by giving agent-mediated agreements a better structure.

The compact is one such structure.

It will need revision. It will need pressure. It will need implementation. But it has to exist before the agent era inherits the web’s broken consent machinery and accelerates it beyond recognition.

The window is open now.

It will not stay open indefinitely.

The work is to make the path real before the alternatives foreclose it.