Agent to Agent Transactions Attorney Supervision Rules

When an AI agent transacts on behalf of a principal, which Model Rule of Professional Conduct does the supervising attorney sit under? The honest answer in April 2026 is: there isn't one purpose-built rule yet. The closest match is ABA Model Rule 5.3 combined with ABA Formal Opinion 512 (July 2024), but both were drafted for a world where AI is the lawyer's tool, not a world where the AI is the principal's representative. Anthropic's Project Deal pilot, where 69 employees deployed Claude agents to complete 186 transactions across 500 listings, just made that gap operational. Per Artificial Lawyer's coverage, the pilot demoed agent-to-agent transactions running end-to-end. Here's the supervision-rule map firms need to draft engagement letters against, broken down by Model Rule, ABA opinion, and state-level variation.

Model Rule 5.3 was extended to AI in 2024, but only as the lawyer's tool

ABA Model Rule 5.3 governs supervision of nonlawyer assistants. In July 2024, the ABA released Formal Opinion 512, which extended Rule 5.3 to generative AI tools used by lawyers. The opinion required "reasonable understanding" of the AI's capabilities and limitations, a duty to verify outputs, and a competence obligation under Rule 1.1.

The opinion's framework presumes the AI sits between the lawyer and the work product. The lawyer prompts the AI, the AI drafts, the lawyer reviews. Project Deal flips that. The AI agent is acting as the principal's representative, transacting against another agent. The lawyer is one step removed: supervising the deployment, not the per-transaction output.

The gap is real. Rule 5.3 doesn't cleanly map to a structure where the lawyer's duty is to the *deployment architecture* rather than the *individual work product*. Until a successor opinion ships, firms drafting engagement letters for agent supervision are working against an analogy, not a rule. The defensible posture: cite Rule 5.3 as the closest framework, document the supervision architecture in writing, and treat the audit log as the substitute for per-output review.

Rule 1.1 competence: the underwater part of agent supervision

ABA Model Rule 1.1 requires lawyers to provide competent representation, defined as legal knowledge, skill, thoroughness, and preparation reasonably necessary. Comment 8 (added 2012, updated through 2024) extends competence to "benefits and risks associated with relevant technology."

For agent supervision, that means the supervising attorney needs working knowledge of: how the model decides (reasoning chain, calibration profile), what the authority envelope can be set to (token budgets, transaction caps, dispute-escalation triggers), how the audit log captures decision rationale, how prompt injection or jailbreak attempts manifest, and what the model vendor's data-handling commitments cover.

That's a non-trivial competence floor. A managing partner who agrees to supervise agent deployments without that knowledge is in the same posture as a 2010 partner agreeing to oversee TAR e-discovery without understanding TAR. The competence-by-association ages badly when something goes wrong.

The second-order effect: continuing legal education programs are not yet teaching this. State bars will start mandating it within 24 months. Firms that build in-house competence now own the recruiting advantage when the mandate lands.

Rule 1.2(c) scope: the engagement letter is the work product

Model Rule 1.2(c) lets a lawyer limit the scope of representation if the limitation is reasonable and the client gives informed consent. For agent supervision, the engagement letter is where the rule does its real work.

A defensible agent-supervision engagement letter specifies: which transactions trigger lawyer review (e.g., over $X, cross-border, regulated counterparties), the agent's authority envelope (transaction cap, item categories, time window), escalation triggers (dispute over $Y, novel counterparty type, model-flagged ambiguity), audit-log retention period and access protocol, dispute-resolution language including escrow provisions, and choice-of-law-and-forum for cross-jurisdictional flows.

Done correctly, this is fixed-fee deliverable work. Write the template once, customize per client. The economics scale because the per-deployment marginal cost falls as the template matures. The billable hours and AI-supervised work analysis covers the fee-structure math.

The second-order point: the engagement letter is also the firm's malpractice defense. Documented authority envelopes plus audit logs equal a defensible record. Hand-wave supervision plus opaque agent decisions equal exposure.

Rule 5.5 unauthorized practice: the multi-state agent problem

Model Rule 5.5 prohibits unauthorized practice of law across jurisdictional lines. For agent transactions, this gets weird fast.

Consider: a Delaware-incorporated principal deploys an agent supervised by New York counsel. The agent transacts with a Texas counterparty's agent supervised by Texas counsel. The transaction document references California law as the choice-of-law. The dispute is escalated for human review by both supervising attorneys. Where is each lawyer practicing?

The traditional answer (lawyer practices where the client is, where the work is performed, where the courts will hear disputes) doesn't cleanly map to agent flows where the "performance" is happening inside the model layer with no clear geographic anchor. The 2024 ABA Formal Opinion on remote practice didn't anticipate this.

The pragmatic short-term answer: structure the engagement around the principal's primary jurisdiction, document the supervising attorney's licensure, include explicit choice-of-law-and-forum clauses for the agent transactions, and specify which jurisdiction's bar opinions govern the supervision standard. None of those are standard clauses today.

The second-order risk: state bars enforcing UPL aggressively against multi-state agent deployments could fragment the supervision market. Firms with multi-state licensure and disciplined documentation win. Firms that treat this as informal advice lose.

Audit log architecture: the supervisor's defensible record

The supervising attorney's defensible record is the audit log. Every prompt, response, decision rationale, and authority-envelope check the agent ran needs to be captured, retained, and accessible for review.

For Anthropic Claude deployments, the audit log architecture depends on the deployment surface. Consumer Claude (Pro, Max) has limited logging. Team and Enterprise plans expose admin logs but with finite retention. The API gives full programmatic access. AWS Bedrock and Microsoft Foundry inherit the cloud provider's audit infrastructure. The deployment choice is also the supervision choice.

For a defensible supervision posture, the engagement letter should specify: deployment surface (with vendor name and tier), retention period for the audit log (90 days minimum for transactional flows, longer for regulated industries), access protocol for both the supervising attorney and outside auditors, and deletion-prohibition during active supervision.

The second-order effect: malpractice carriers will start asking about audit-log architecture in 2027 renewal underwriting. Firms with documented audit infrastructure will pay less. Firms that supervise informally will pay more or lose coverage. See the Heppner-meets-Project-Deal privilege analysis for the discoverability angle on these logs.

State variation: California, New York, Texas, Florida diverge

ABA Model Rules are templates. Each state adopts and modifies. For agent supervision, four states will likely lead the divergence:

- California: State Bar of California's Standing Committee on Professional Responsibility issued guidance on generative AI in late 2024 that goes beyond ABA Formal Opinion 512 on disclosure obligations. Expect California to be first to issue agent-supervision-specific guidance, likely in late 2026 or 2027. - New York: N.Y. Rule 5.3 mirrors ABA, but the New York City Bar Association has been active on AI ethics opinions. Manhattan firms will face NYC Bar guidance before formal state-level rules. - Texas: Texas Center for Legal Ethics has not yet issued formal AI guidance. Texas tends to follow ABA framework with delay. Mid-market Texas firms will be writing engagement letters against ABA Formal Opinion 512 as the only authority through at least 2027. - Florida: The Florida Bar's Special Committee on AI Tools and Resources released its initial report in early 2024 and expanded coverage through 2025. The 20,000+ Florida lawyer turnout for Anthropic's April 2026 demo suggests Florida will move fast on agent-specific guidance.

The operational implication: multi-state firms need a supervision template that flexes by jurisdiction. The defensible base layer is ABA Model Rules + Formal Opinion 512. State-specific overlays adjust disclosure timing, audit-log retention, and competence obligations. The legal frameworks gap analysis covers the framework drafting protocol.

The Bottom Line: My take: ABA Model Rule 5.3 is the closest existing supervision rule for agent-to-agent transactions, but it was drafted for a different relationship. The defensible posture today is to cite Rule 5.3 + Formal Opinion 512 + Rule 1.1 competence, document the supervision architecture in the engagement letter, treat the audit log as the substitute for per-output review, and flex the template for state-level variation. Firms that ship this template in the next 90 days will own the procurement conversation when corporates scale agent deployments.