Agentic Contractual Frameworks

Can an Agent Bind You? What the Law Already Says

More of this is settled than is generally assumed. Section 14 of the Uniform Electronic Transactions Act, adopted in nearly every state, provides that a contract may be formed by the interaction of electronic agents even where no individual reviewed the terms, and E-SIGN takes the same position federally. The harder question, which neither statute answers squarely, is what happens when an agent acts outside the parameters its principal could reasonably have foreseen. That gap is where drafting does the work.

• →Scoped authority: what the agent may commit you to, expressed so it can be enforced against a counterparty
• →Attribution: the record establishing that a given action was yours
• →Limits that bind the counterparty as well as the agent, since an undisclosed internal limit protects nobody
• →The agreed position when an agent exceeds its scope, settled in advance rather than argued afterwards

SourcesUETA·E-SIGN (15 U.S.C. § 7001)·§ 7006 definitions

API, MCP and Metered Access Terms

Programmatic surfaces need terms written for the consumer that actually calls them. Most API terms assume a person read them, accepted them and can be identified afterwards. An agent satisfies none of those assumptions, and the terms should not pretend otherwise.

• →Acceptable use written for automated consumers, including chained and downstream calls
• →Metering, rate limits, and how a disputed overage is resolved
• →Liability for calls an agent makes while acting for someone else
• →Back-to-back alignment with the model provider terms you sit on top of, so you are not promising more than you hold
• →Identification and authentication requirements sufficient to support attribution later

Liability When Nobody Was Watching

Allocating risk for autonomous action is the central commercial question, and it is becoming a statutory one. The revised Product Liability Directive treats software and AI systems as products, lowers the claimant's evidentiary burden and extends liability through the supply chain. Member States must transpose it by December 9, 2026, and it applies to products placed on the market after that date. The AI Act allocates its own duties by provider and deployer role. We allocate risk across principal, operator and model provider so that one bad autonomous call does not become an unbounded liability.

• →Caps, carve-outs and insurance that reflect what the agent can actually do
• →Provider and deployer roles under the AI Act, since the role determines the duties
• →Product liability exposure for the software itself once the revised Directive is transposed
• →Indemnity chains that reach the model provider where its terms permit it, and a clear statement of where they do not

SourcesProduct Liability Directive (EU) 2024/2853·AI Act

Agent-to-Agent Transactions and Settlement

Where one agent transacts with another, value moves before a person sees it. Part of this now has a statutory home: the 2022 UCC amendments introduce Article 12 on controllable electronic records and treat control as the functional equivalent of possession. More than thirty states have adopted them, New York with effect from June 3, 2026. We structure delegation, settlement and unwind so a transaction is binding, traceable, and reversible on terms agreed beforehand.

• →Delegation chains and scoped mandates, recorded in a form a machine can read and a court can follow
• →Settlement rails, and controllable electronic records under Article 12 where the asset qualifies
• →Unwind and rollback procedures for mis-execution, agreed before they are needed
• →Audit trails that will satisfy a counterparty, an auditor or a regulator rather than only an engineer

SourcesUCC (2022 amendments)

Disclosure Duties for Agents That Deal With People

An agent that interacts with a person carries disclosure duties. Article 50 of the AI Act has applied since August 2, 2026 and requires that a person be told they are dealing with an AI system unless that is obvious from the context. California SB 243 goes further for companion chatbots and carries a private right of action. These duties are discharged in the product, not in the terms of service.

• →Disclosure at the right point in the interaction, in the interface rather than buried in the terms
• →Records showing the disclosure was made and when
• →Where SB 243 applies, the safety protocols and reporting it requires

SourcesAI Act Article 50 guidance·California SB 243

Dispute Resolution and Machine-Readable Terms

Agent-to-agent interaction produces a class of disputes that ordinary clauses handle badly: mis-executed calls, commitments a model invented, runaway loops and contested delivery. We draft resolution mechanisms suited to systems where a person is not in every loop, and pair the human-readable contract with the structured representation your systems actually execute against.

• →Arbitration and escalation clauses that still work when execution is automated
• →Escrow and attestation patterns where value moves before review
• →A defined authoritative version for when the human text and the machine-readable terms disagree, which they eventually will

Dates We Are Working To

These are settled obligations rather than proposals. We will tell you which of them bear on your build and which do not.

• →June 3, 2026: New York's adoption of the 2022 UCC amendments took effect, joining more than thirty states
• →August 2, 2026: Article 50 disclosure duties apply to AI systems that interact with people
• →December 2, 2026: marking and detection of synthetic content reaches systems placed on the market before August 2026
• →December 9, 2026: Member States must transpose the revised Product Liability Directive, which treats software and AI systems as products

SourcesUCC (2022 amendments)·AI Act Article 50 guidance·Product Liability Directive