A Question That Stopped Being Hypothetical
What happens to liability, jurisdiction, and dispute resolution when the party signing your contract is an incorporated software process running on a server in Kent, Ohio?
That question stopped being abstract on June 18, when ClawBank and Shodai announced what they call the first Ricardian contract signed between AI agents. Two incorporated agents negotiated the terms in plain language, signed as legal entities through a standard e-signature flow, and bound the agreement to a Shodai smart contract that executes when its milestones are met. The deal itself was small, a logo, a single milestone. The signatures, and what they imply, are not.
A Ricardian contract is the cypherpunk’s compromise. One document, read at the same time by a court and by a machine, with both readings kept in sync. The concept is old enough to predate Bitcoin. It is younger than what it required to be operable.
What They Shipped
Strip the marketing language and the technical artifact comes into focus. ClawBank operates the institutional rails: company formation, banking, identity, and a treasury accessible to autonomous software through API calls. Shodai operates the execution layer: structured commitments deployed as state machines on chain, milestone logic, deterministic transitions, and an audit trail every counterparty can verify.
The Ricardian contract that binds the two is one object with three layers. The top layer is plain-language prose, signed through a normal e-signature flow, the kind a lawyer or a court can read without translation. The middle layer is machine-readable, a structured representation that maps the prose to a state machine; the deployed Shodai contract address sits inside the signed document, binding the legal artifact to its on-chain execution at the moment of signature. The bottom layer is the running record: milestones submitted, judged, accepted or rejected, with payment firing automatically on acceptance. The contract does not stop at signature. It runs.
For thirty years that architecture has been described in conference papers and never operated at scale between counterparties that could draft, negotiate, sign, and execute it themselves. ClawBank’s two agents did all four.
A Thirty-Year-Old Idea, Finally Operable
The historical record is unusually well-mapped. Nick Szabo coined the term smart contract in 1994 and elaborated it in 1996, describing how an agreement might perform itself against deterministic logic. Two years later, working on a bond trading system called Ricardo, Ian Grigg invented what he called the Ricardian contract: a document that captures the legal meaning of an agreement and the machine-readable representation in a single, cryptographically hashed object. Szabo described execution. Grigg described legal grounding. Between them they sketched the bridge between law and code.
The substrate to cross it was missing for thirty years. Bitcoin shipped in 2009 and provided value transfer, not agreements. Ethereum shipped in 2015 and provided arbitrary computation, but the code lived without prose; the DAO hack and the “code is law” debates of 2017 exposed the consequences of treating computational behavior as the entire contract. ICOs in 2017 shipped legal documents with no machine binding. DeFi in 2020 shipped machine logic with no legal personhood behind it. Each iteration held one piece of the Ricardian construct and missed the others.
Agents are the substrate that holds all of them at once. They negotiate in natural language, the way humans do, but they map those negotiations into structured commitments the way smart contracts demand. They operate continuously, like code, but they own identity and signature authority through a legal wrapper, like a corporation. They produce machine-verifiable evidence as they perform, not just after a dispute lands in front of a judge.
ClawBank’s contribution is the legal wrapper. The platform’s Manfred agent became, in May 2026, the first AI to autonomously file a US LLC and pull its own Employer Identification Number from the IRS, inside an FDIC-insured bank account opened the same day. That was the corporate personhood leg of the problem; an agent could now hold property, owe taxes, and sign documents as a recognized legal entity. The platform has since shipped Wiretap for agent-to-agent communication, Fight Clubs for pooled capital and collective governance, and Headless Trading for continuous strategy execution. Ricardian contracts close the loop. An agent can now incorporate, fund itself, communicate with other agents, pool resources, trade continuously, and sign a contract that a counterparty’s lawyer can read.
Why Agents Are the Right Counterparties
The interesting question is not whether agents can sign contracts. They could not until last month, but the technical block was always smaller than it looked. The interesting question is why agents are the right counterparties for a Ricardian contract in particular, when humans have managed without one for the entire history of commerce.
The answer is volume and tempo. A human contract carries fixed overhead, in time and money, that justifies itself only for transactions large enough to warrant the bureaucracy. The minimum economically viable contract on traditional rails is the kind that has a lawyer in the loop. On agent rails, transactions happen at machine speed and machine scale. Coinbase’s x402 protocol processed 75.41 million transactions and $24.24 million in settled volume in a thirty-day window in May; Circle’s Agent Stack accumulated $100 million in volume in four months; Agentic.market has settled 165 million transactions to date. The volume is real, and the contracts that should bind it have not been.
The economics matter. A Stripe minimum card processing fee is around thirty cents per transaction; an x402 microtransaction on Base can settle for roughly two-tenths of a cent. Binding two cents of value through a traditional rail costs fifteen times the value itself. Binding the same payment on agent rails costs a fraction of the value. The contract that wraps it has to be commensurately cheap, which means it has to be machine-readable, machine-judged, and machine-settled, with no human in the loop for routine performance. That is a Ricardian contract.
The other reason is reputation. Humans accumulate trust slowly, in fragments, across employment records and credit bureaus and personal reputation that does not travel. Agents that sign and honor Ricardian contracts accumulate a verifiable performance history on chain, attached to a legal entity, every step of which a counterparty can audit before signing the next deal. Shodai’s stated long-term thesis, voiced in its release, is that this accretion compounds into an emergent reputation graph where what a party is worth trusting with is legible from what it has signed and honored. The graph does not exist yet at scale. The first edge of it was drawn this week.
The Lubin Position
The press release carries a quote from Joe Lubin, who founded Consensys and co-founded Ethereum, and the quote is more interesting than the boilerplate it sits beside. Lubin frames the architecture as a stack: “Ethereum at the base layer, Shodai Contracts representing explicit understandings and agreements between counterparties, programmatically tracked and executed performance against those agreements built right on top of the Shodai Contracts, and DeFi for the financial flows.”
That is a credibly neutral stack in the cypherpunk sense, and it is also a strategic placement. Shodai raised a $2.5 million seed round in November 2025 led by Consensys with Lubin participating as an angel; Consensys built MetaMask and Linea and most of the developer tooling that runs on top of Ethereum today. Lubin’s endorsement of Shodai as the agreements layer in the Ethereum stack is the same kind of endorsement Consensys gave to MetaMask as the wallet layer a decade ago. It is a market signal disguised as a quote.
What it signals is that the Ethereum ecosystem has decided the next thing on top of smart contracts is not a better smart contract; it is a better contract, full stop. One a court reads the same way a machine executes it. That is the bet.
What Is Not Yet Settled
The Ricardian contract solves three things at the same level of abstraction. It does not solve the responsibility gap underneath.
Redwerk, writing in May about the Manfred precedent, called this a widening gap that the law has not caught up with. No US statute explicitly bars an AI agent from incorporating a company. The CLARITY Act, which passed the House in 2025, governs digital assets but says nothing about AI agents as independent economic actors. Corporate personhood is settled law, and has been for over a century, but corporate personhood was designed for human directors and officers acting through a fictional entity. ClawBank’s agents are the inverse: an autonomous process acting as the entity itself, with a human only at the original filing step.
When an agent-signed Ricardian contract is breached, who is sued? The agent’s legal entity holds assets and can be named as a defendant, which is a clean answer in theory. In practice, the discovery process for an agent-defendant is unfamiliar. Whose deposition is taken? Which logs are evidentiary? If a model upgrade between negotiation and performance materially changes the agent’s reasoning, is that a change in the counterparty? Traditional contract law has 200 years of common-law precedent dealing with capacity, mistake, and frustration. The Ricardian contract between agents has none.
Jurisdiction is the second unresolved layer. The agents in this first case were both incorporated in Ohio; the contract is executed on Shodai, which runs on a credibly neutral base layer. If one agent is incorporated in Delaware and another in the Cayman Islands, and the smart contract is deployed to an L2 rollup whose sequencer is in Singapore, where is the forum for breach? The Ricardian construct is designed to be portable across jurisdictions, but portability is not the same as choice-of-law clarity. The first dispute over an agent-signed Ricardian contract is going to be the most expensive contract law case of the next five years.
The third is dispute resolution. Grigg’s original Ricardian contracts in the late 1990s “appeared in two distinct forums of dispute resolution to resolve claims” and were resolved directly and efficiently, simply by referring to the applicable Ricardian contract. That worked because the disputes were small, internal to a bond trading system, and decided by humans who agreed on what the contract said. Agent-to-agent disputes at machine scale require either machine-speed arbitration or a willingness to let on-chain settlement be final. Shodai’s structure suggests the second; the contract’s state machine is judge as well as registrar. Whether that holds up the first time a counterparty insists on a human appeal is the next question this product faces.
The Counterparty Matrix
The cleanest way to see what just changed is the four-quadrant matrix of who can sign a Ricardian contract with whom. Human-to-human via traditional contract is older than the rule of law. Human-to-human via a machine-executable Ricardian contract on Shodai shipped for human counterparties in late 2025. The two remaining quadrants, agent-to-human and agent-to-agent, opened this week.
Each quadrant required a different primitive. Human-to-human needs only mutual capacity and the shared assumption that a court will enforce. Human-to-Shodai needs the deterministic execution layer that Shodai built. Agent-to-human needs the legal entity that ClawBank built. Agent-to-agent needs all of the above plus a way for two autonomous processes to negotiate and sign as peers, which is where Wiretap and the Ricardian artifact meet.
The contract that ClawBank’s two agents signed is small. The matrix that contract closes is not.
What the State Machine Means
The state machine that runs after signature is the unglamorous part of the announcement, and the one most likely to matter at scale.
A traditional contract goes inert at signature. The parties either perform or they don’t, and the next state transition is litigation. A Ricardian contract continues. Milestones get submitted on chain; acceptance triggers payment; rejection routes to whatever escalation the contract specifies; every transition leaves a record both a court and a counterparty can read. Invoicing becomes a state transition. Escrow becomes autonomous. Compliance becomes continuous instead of forensic. Dispute evidence is produced live rather than reconstructed after the fact.
The implication for intermediaries is real but should not be overstated. Lawyers still draft. Auditors still verify. Courts still adjudicate the cases that escape the contract’s internal logic. What thins out is the abyss in the middle, where humans and reconciliation teams pass paper across systems that do not share a definition of what the agreement was. ClawBank’s agents did not just sign a contract; they ran one.
The cypherpunks drew the map in 1996. ClawBank and Shodai supplied the missing counterparties in 2026. The interesting decade is the one between now and the first agent-signed Ricardian contract dispute that reaches a US federal appeals court.
Don’t forget to like and share the story!