A Senate Bill on AI Agents Sets Conditions on the Buyer

← Back · ACT III · Chapter 17: VIGOR: Governance and Realization

On 29 June 2026, Senator Mark Warner released a discussion draft called the Artificial Intelligence Access, Gatekeeper Exchange, and Nondiscriminatory Transfer Act, and on 21 July he introduced it in the Senate as S. 5051. Most of the coverage read it as a platform-access story: which AI agents get to plug into the large consumer platforms, and on what competitive terms. That is in the bill. It is not the part that matters to a buyer.

The bill is the first serious federal attempt to define what it means for software to act on a person’s behalf, and the answer it reaches is not a code of conduct. It is a list of conditions the arrangement has to have, written so that the duties cannot be waived.

What the Text Actually Requires

The bill creates a category it calls the custodial user agent: software a person authorizes to act for them online. Section 3 defines that authority as transparent, documented, limited and revocable. Section 3(g) sets the duties. The agent may not take actions that benefit itself to the user’s detriment, or that run against the user’s directions or reasonable expectations. It must act with the care, skill and diligence an ordinarily prudent person would exercise in the same position. It must keep real-time records of what it does on the user’s behalf and provide them to the user, unless the user has told it to delete them. It may not hand its authority to another system without the user’s express, specific and revocable consent. Section 3(g)(3) then says none of these duties may be waived, limited or modified by contract, by terms of service or by any form of user consent.

Nothing in that list asks the agent to be well intentioned. There is no ethics pledge and no certification of character. Certification does appear in the bill, as registration with the Federal Trade Commission and, under section 3(d)(4), periodic assessment by an independent certification body, rather than as a claim about conduct. The text specifies what the arrangement must have no matter who operates it, and then puts the core duties beyond the reach of a contract so nobody can turn a condition back into a preference.

A Condition Holds When Nobody Is Watching

This is the distinction most governance programs get backwards. A behavior is something a person or a system performs, which means it stops the moment attention moves elsewhere. A condition is something the structure has. The authority is documented and revocable from the moment it is granted, the records run whether or not anyone is reading them, and accountability sits with whoever authorized the work rather than with whoever ended up holding the task.

Congress did not arrive there out of pessimism about software. It arrived there because delegated authority is the one arrangement where good intentions are unverifiable at the moment they matter, and a record is the only thing that answers afterwards.

The Enterprise Has No Equivalent

Turn the bill inward. Inside most companies, agents are already acting on delegated authority: on somebody’s credentials, against somebody’s data, producing work somebody will claim as a result. Ask of any one of them the three questions the bill answers for consumers. Who authorized this agent, and is the authorization documented and bounded? What has it done, and can anyone outside the deploying function read the record? Can its authority be revoked today, cleanly, by someone other than the person who set it up?

Most operating models cannot answer the third, and a good many cannot answer the first. That has been noticed publicly, from the compliance side. Sanchit Vir Gogia of Greyhound Research told CIO on 30 June that a right to revoke means very little until the enterprise can say what is being revoked, from whom, and across which systems. Forrester’s Biswajeet Mahapatra said in the same piece that tying an agent to its authorizing user creates a continuous traceability requirement. Both are describing a gap that is older than the agents. The function that deploys the tool and the enterprise that answers for the result have never been bound to the same record. Agent traffic makes that visible at machine speed.

What Waiting Costs a Buyer

Warner also wrote to Treasury Secretary Scott Bessent, ahead of a hearing, asking Treasury to work out how existing law allocates liability when an agentic system causes a loss, and arguing that an agent should owe a duty of loyalty to the principal it acts for. The direction is legible: identity bound to an accountable operator, records by default, authority revocable by default, duties that survive delegation.

A buyer can wait and inherit that structure from statute, on someone else’s definitions and someone else’s schedule. Waiting costs more than the compliance work at the end of it. Every dollar of agent spend made in the meantime buys a result nobody can trace back to the authority that permitted it, which is the same condition that leaves a purchase claiming a return no record supports. Ownership is a claim until there is a record, and delegated authority is a claim until the record names who granted it.

Custody is a condition a structure has on the day the agent acts. Congress has now written down what that condition consists of, in a text any buyer can read before anyone is required to.

References: S. 5051, Artificial Intelligence Access, Gatekeeper Exchange, and Nondiscriminatory Transfer Act of 2026 (the AI AGENT Act of 2026), introduced in the Senate 21 July 2026 and referred to the Committee on Commerce, Science, and Transportation, text at govinfo.gov; discussion draft and section-by-section summary, Office of Senator Mark Warner, released 29 June 2026. Prasanth Aby Thomas, “How the Senate’s AI AGENT Act could reshape enterprise AI governance,” CIO, 30 June 2026. “Ahead of Hearing with Secretary Bessent, Warner Encourages Treasury to Better Understand and Develop Rules Around Agentic AI,” Office of Senator Mark Warner.

Keep reading: ACT III · Chapter 17: VIGOR: Governance and Realization