The agent can prove its name. Nobody is vouching for it yet.

Proving an AI agent's name is nearly solved. Deciding whether to let it in is up for sale.

· 5 min read

On Sunday evening, 20 September, people who'd sent Meta's new Muse assistant to shop on Amazon got a legal notice where the cart should have been. "Continued access by an unauthorized AI agent violates Amazon's Conditions of Use, to which our customers have agreed," the popup read. Amazon's spokesperson framed it as manners: shopping apps "should operate openly and respect service provider decisions about whether or not to participate."

Since May 2025 Amazon's Conditions of Use have spelled out what openly means: an agent announces itself in its user-agent string, the line of text a browser sends with each request, as "Agent/agent name". The next day, Amazon amended its lawsuit against Perplexity, whose Comet agent sends Google Chrome's string instead. "Perplexity chooses that value, Perplexity codes it into the software Perplexity distributes," the filing says.

One demand: say your name. It's the half of the problem the industry has nearly solved. An agent can now declare who it is, and the company that runs it can sign for it (using something called Web Bot Auth). Almost nobody can yet tell a store whether to let it in, and the people offering to are the ones who already keep score on you.

The court won't sort them

On 4 August the Ninth Circuit declined to do the sorting. The panel held that the shopper is the one who reaches Amazon's servers: "It is the user who 'accesses' Amazon's computers, with the help of the Assistant." And of the agent: "However advanced the Assistant currently is, it is a tool, not a person for statutory purposes." On 10 September the full court refused to rehear the case, and no judge asked for a vote.

The ruling is narrow, and says so: "We do not establish a new legal regime governing agentic AI." Stores can still refuse agents, and Amazon is leaning on its own terms: the Muse popup cites its Conditions of Use and accuses nobody of hacking. Its amended complaint still pleads computer fraud, now with new claims about Comet on the iPhone. On this record, the federal anti-hacking law isn't what separates good agents from bad ones.

Amazon told the court the decision "degraded website owners' ability to set the terms" for AI agents. The panel weighed the other side: an injunction "would impair consumer choice and needlessly limit development of a nascent technology." Amazon's taste for openness is also a business position, since its ad revenue depends on people looking at its pages.

That leaves a store its terms of service and whatever the agent can prove about itself.

The passport is nearly built

What an agent can prove is its name. Cloudflare says operators including OpenAI, Google and AWS sign their agents' requests with Web Bot Auth, and counts "more than 500 billion verified bot requests each week" on its own network. An IETF working group published a standards-track draft of the scheme on 1 September, from Cloudflare and Google engineers. Visa launched its Trusted Agent Protocol with Cloudflare in October 2025, so merchants can "securely recognize and transact with Visa-approved AI agents," and describes it as "aligned with" the same effort.

A second set of tools ties the agent to a person. The FIDO Alliance, the body behind passkeys, opened work in April on agent sign-in and payments, starting from Google's AP2 and Mastercard's Verifiable Intent.

World sells its agent tools as "a power of attorney for your agent." Vercel's Passport, built for a company's own agents and apps, hands an agent a signed token naming the person who signed in. Identity plus authority is what a passport carries: who you are, and who sent you.

None of it is finished. "Trust frameworks for agent-driven interactions do not yet exist," FIDO says. The IETF group's charter aimed its standards for April 2026, and its draft is only on its second working-group version, -01, with the usual rounds of revision ahead. The webbotauth group meets for two hours at the IETF's San Francisco meeting on 17 November.

A passport doesn't vouch

The engineers writing the standard draw the line themselves. A valid signature, their draft says, "says nothing about who operates the Agent, whether the Agent is benign, or whether the request is authorized."

The draft "does not define how trust is accrued, held, or exchanged." The charter rules out "Tracking or assigning reputation to particular bots," along with authenticating the end user. Sensible for a protocol, and it leaves the hardest question with no owner.

Cloudflare's documentation shows where it bites. It labels every verified agent Direct or Intermediary, and an intermediary is "an agentic service that a wide range of end users can operate." Cloudflare's docs spell out the limit: "you may trust the intermediary operator, but not necessarily every end user driving it." The store knows the brand on the passport and nothing about the traveler.

A name, declared or signed, does give reputation something to stick to. In August 2025 Cloudflare de-listed Perplexity as a verified bot after watching it use "a generic browser intended to impersonate Google Chrome on macOS." That was a reference withdrawn, by one company, against a known name, and it's about as far as the system goes. The one agent-reputation proposal I found at the IETF is an individual draft from one company, with "no formal standing."

The referee is a credit bureau

The reference is arriving as a product. On 30 April Experian announced Agent Trust with Visa, Cloudflare and Skyfire.

It binds each agent to a verified person and promises "a continuous, auditable history between who the user is, and how the agent behaves over time." Its Agent Registry maintains "dynamic trust scoring for AI agents based on behavior and other risk signals." Akamai, working with Experian and Skyfire on what they call Know Your Agent, links agents "to authorized human users."

Follow that design to its end and the agent's reputation is yours. A score built from how your agent behaves, filed against your verified identity by a company that already keeps credit files, is a credit check with a new input. Experian's chief innovation officer is cheerful about the lineage: "We already help define trust in financial transactions; now we're bringing that same leadership to agentic commerce."

Whether any of it works, I can't tell from outside. The launch gave no adoption figures and no scoring model, so nobody can say how the score is made, who reads it or how you'd dispute one. A private referee brings a credit bureau's familiar problems: an opaque model, disputes, people scored out without knowing why. The security analyst David Shipley's line about World, "Private sector control of personhood feels Hollywood-style cyber dystopian," fits a referee of agent conduct too.

Once every agent carries a passport, the scarce part is the reference: a record, kept by someone, of how the person behind the agent has behaved. For now the company offering to keep it already keeps your credit file.

Work together

Have a project in mind?

I help purpose-driven organizations build the technology they need to tell their stories. If that sounds like you, let's talk.

Get in touch with Five59 Labs →

The newsletter

Get the receipts.

Every other Friday, on how trust gets built and broken in a synthetic world. Written by someone who builds with these tools daily, and shows the receipts.

No spam. Unsubscribe any time.

Work together

Have a project in mind?

I help purpose-driven organizations build the technology they need to tell their stories. If that sounds like you, let's talk.

Get in touch with Five59 Labs →