Legal Frameworks Needed AI Agent Marketplaces 2026

Legal IT Insider's April 27, 2026 headline on Anthropic's Project Deal pilot named the gap directly: "the legal frameworks don't exist." After Anthropic ran 69 employees, $100 budgets, and 186 completed agent-mediated transactions, the operational risk became measurable. The framework gap spans agency law, contract law, antitrust, professional ethics, evidence, and dispute resolution. Each gap is a procurement risk for corporates running pilots and an opportunity for firms that ship defensible templates first. Here's the framework drafting protocol: what's missing, where the analogs sit, and what firms should write before the second pilot ships.

Agency law: Restatement (Third) needs an agent-as-software section

Restatement (Third) of Agency § 1.01 defines an agent as "a person who agrees to act on behalf of another." Comments contemplate human or corporate agents. The duties enumerated in §§ 8.01-8.11 (loyalty, care, confidentiality, good conduct, no undisclosed material benefit) apply to "agents" as defined.

The gap: software agents don't fit the § 1.01 definition cleanly, but the duties have to flow somewhere. The defensible reading allocates duties to the principal, the human or corporate entity that deployed the agent, and treats the AI as an instrumentality. That allocation works for first-pilot scenarios. It strains under multi-party flows, dual representation, and conflicts.

What the framework needs: an explicit Restatement section addressing software agents, allocating duties between the principal, the supervising attorney, and the model vendor. Restatement supplements take years. State courts can fill the gap faster through case-by-case rulings, but no appellate court has reached the question on Project Deal-style facts.

The operator move: firms drafting engagement letters now should explicitly allocate duties in writing: principal bears agent-side duties, supervising attorney bears Model Rule 5.3-style supervision duties, model vendor bears product-liability duties for design defects. Document the allocation. The AI agent fiduciary duty analysis covers the duty stack.

Contract law: UCC Article 2 and the formation gap

UCC § 2-204 lets contracts form "in any manner sufficient to show agreement." That's flexible enough to accommodate agent-mediated formation. The gap is enforceability when the agent goes off-script.

If an agent forms a contract beyond the principal's intended authority envelope, several questions follow: is the contract void, voidable, or enforceable against the principal? Does apparent authority apply when the counterparty couldn't reasonably know the agent's limits? Does the model vendor bear any responsibility for the agent's behavior?

UCC Article 2 doesn't answer these directly. The closest analogs are corporate authority cases (where corporate agents bind the corporation under apparent authority) and electronic signature cases (where automated systems can form contracts under E-SIGN and UETA). Neither was drafted with software agents reasoning about counterparty positions in mind.

The pragmatic short-term answer: explicit authority envelope clauses in transaction documents. The principal's agent operates within disclosed authority. Counterparties acknowledge the disclosed authority. Transactions outside authority require human escalation. The escrow and arbitration framework covers the dispute-resolution language for off-script transactions.

The second-order reform need: NCCUSL or ALI-led Restatement supplements specifically addressing software agent contract formation. That's a multi-year process. Firms drafting commercial contracts now are working without authoritative guidance.

Antitrust: algorithmic collusion when models share weights

Sherman Act § 1 prohibits agreements in restraint of trade. Cartel cases require proof of agreement. The textbook frame is human agents on phone calls.

Agent-to-agent transactions raise a different question. When two Claude instances negotiate against each other and reach predictable outcomes because they share model architecture, is that "agreement" under Section 1? The FTC has flagged algorithmic collusion concerns in prior guidance. The DOJ Antitrust Division has not yet brought a Section 1 case on agent-to-agent facts.

Three possible enforcement frames will compete:

- Tool-of-agreement frame. Shared model weights are a means of facilitating coordinated behavior, requiring disclosure and remedies analogous to information-sharing cartels. - Parallel-conduct frame. Predictable agent behavior is parallel conduct without explicit agreement, requiring "plus factors" to elevate to a Section 1 violation. - New-category frame. Agent-mediated coordination is structurally distinct from human collusion and requires statutory or regulatory response rather than common-law extension.

None of the three is clearly the right answer. The framework gap creates enforcement uncertainty. Corporate counsel deploying agents need to anticipate procurement requirements within 18 months: counterparty-detection features (flagging when the counterparty agent runs the same model), audit-log retention sufficient for antitrust compliance review, and escalation protocols for transactions in concentrated markets.

Professional ethics: ABA Formal Opinion 512 needs an agent-supervision successor

ABA Formal Opinion 512 (July 2024) extended Model Rule 5.3 supervision duties to generative AI as a tool used by lawyers. The opinion presumes the AI sits between the lawyer and the work product.

Project Deal flips that relationship. The AI agent acts as the principal's representative, transacting against another agent. The lawyer is one step removed, supervising the deployment architecture rather than per-transaction output. ABA Formal Opinion 512 doesn't cleanly address that posture.

A successor opinion needs to address: how Rule 5.3 supervision applies when the AI is not the lawyer's tool but the client's representative; how Rule 1.1 competence flexes for agent-deployment knowledge (token budgets, calibration profiles, prompt-injection risk); how Rule 1.2(c) scope-limitation language should specify the agent's authority envelope; how Rule 5.5 unauthorized practice applies to multi-state agent transactions; and how Rule 1.6 confidentiality interacts with the Heppner privilege gap.

The ABA Formal Opinion process moves slowly. State bars can issue parallel guidance faster: Florida, California, and New York are likely first. The agent supervision rules deep-dive covers the state-by-state map for engagement-letter drafting.

Evidence: Federal Rules don't yet address agent transcript authentication

Federal Rules of Evidence 901 and 902 govern authentication. FRE 902 lists self-authenticating documents: categories that don't require extrinsic authentication evidence. The 2017 amendments added FRE 902(13) (electronic records by qualified person certification) and 902(14) (data copied from electronic devices by qualified person certification).

The gap: agent transaction transcripts (prompts, responses, decision rationales) are electronic records that don't fit cleanly into FRE 902(13) or 902(14). They're not records of "electronic records" in the traditional sense; they're records of agent reasoning. Authentication requires showing the transcript is what it purports to be, but the underlying generation process (model inference) doesn't have the deterministic chain-of-custody that traditional electronic records have.

The second-order issue: chain-of-custody for agent transcripts depends on the deployment surface. Consumer Claude has limited audit infrastructure. Enterprise tiers have stronger logging. AWS Bedrock and Microsoft Foundry inherit cloud-provider authentication. Litigation involving agent transcripts will fight over which deployment surface produced the record and what authentication that surface enables.

The framework reform need: FRE amendment specifically addressing AI agent record authentication, or ALI principles guiding judicial discretion under FRE 901. Both will take years. In the meantime, firms drafting engagement letters should specify deployment surface, audit-log architecture, and authentication protocols in writing; the documentation itself becomes the authentication foundation.

Dispute resolution: AAA, JAMS, and the Federal Arbitration Act presume humans

The Federal Arbitration Act (9 U.S.C. § 1 et seq.) presumes human signatories. Section 2's "written provision in any maritime transaction or a contract evidencing a transaction" contemplates human-executed contracts. Sections 3 and 4 (stay of proceedings, compelling arbitration) presume human parties.

For agent-mediated transactions, the dispute resolution analysis hits multiple gaps: who is the "party" to the arbitration agreement (the principal, or the agent acting as principal's representative); whether the agent's assent constitutes the principal's assent; whether arbitration provisions in transaction documents formed by agents are enforceable; and what arbitral forum has jurisdiction when the transaction was formed in the model layer.

AAA and JAMS rules don't explicitly address agent-formed agreements. International arbitration frameworks (UNCITRAL Model Law, New York Convention) similarly presume human or corporate parties.

The pragmatic framework: explicit choice-of-forum-and-law clauses in agent-deployment authority envelopes. Specify that the principal is the party to any arbitration. Specify the arbitral forum. Specify the substantive law. Document the principal's consent to be bound by agent-formed arbitration agreements. None of these are standard clauses today.

The escrow and arbitration AI agent framework covers the clause stack and the procedural architecture for human escalation when agent-resolution fails.

What firms should write before the second pilot ships

The framework gap is real and persistent. Firms that ship defensible templates in the next 90 days will own the procurement conversation when Project Deal-style pilots scale.

The template stack should include:

1. Master engagement letter for agent-supervision retainer. Specifies which transactions trigger lawyer review, the agent's authority envelope, escalation triggers, audit-log retention and access, and choice-of-law-and-forum. 2. Per-deployment authority envelope document. Negotiable per client: transaction caps, item categories, time windows, dispute thresholds, counterparty restrictions. 3. Audit-log architecture specification. Deployment surface, retention period, access protocol, deletion-prohibition during active supervision. 4. Disclosure language for transaction documents. Counterparty acknowledgment of agent representation, deployment surface, model vendor, and authority envelope. 5. Dispute-resolution clause stack. Escrow language, escalation triggers, arbitration forum, choice of law, human-checkpoint protocols. 6. Antitrust compliance protocol. Counterparty-detection requirements, audit-log retention sufficient for review, escalation in concentrated markets. 7. Privilege-and-confidentiality posture. Deployment surface decisions, retention and access protocols, segregation of privileged from non-privileged transcripts.

Done correctly, this is a fixed-fee productized practice. Write the template once, customize per client. The first firm to publish a defensible whitepaper plus engagement letter template plus state-by-state supervision analysis owns the next decade of inbound from corporates running these pilots.

The Bottom Line: My take: The framework gap spans agency, contract, antitrust, ethics, evidence, and dispute resolution. Each gap is real, persistent, and unlikely to close through formal reform on a timeline that matches pilot adoption. The defensible posture is to draft the template stack against the gaps directly: engagement letter, authority envelope, audit-log architecture, disclosure language, dispute-resolution clauses, antitrust protocol, and privilege posture. Firms that ship this template in 90 days own the procurement conversation.