Agents aren’t humans (but they kinda are)

Cloudflare has announced the formation of the Trusted Agent Delegation (TAD) coalition, a cross-industry group meant to work through how websites should recognize, serve and hold accountable the people behind remotely hosted AI agents. The announcement came in a Cloudflare AI post that also lays out the company's position on agent identity and its preferred standards. The coalition's stated goal is to gather representatives from across the ecosystem, review existing protocols, sharpen the problem statements, and take well-defined problems to the IETF, with side meetings at IETF events and workshops planned for later. Cloudflare is inviting agent and agent-runtime builders, edge and cloud platform operators, and teams running commercial or compliance infrastructure to join.
Cloudflare's core argument is that a personal AI agent — an automation that takes a human's instructions and pursues a goal by driving a cloud computer or browser, typically run by a separate operator — is best understood as an intermediary. Sites have dealt with intermediaries before, from cloud browsers to proxies and VPNs, but agents differ in scale: they do not tire, run continuously, operate in parallel and tolerate friction no human would. The posted examples of such agents are Meta's Muse and OpenAI's dots. Much of the web's capacity planning and monetization assumes human limits, the post notes.
The problem, in Cloudflare's framing, is one of visibility rather than intent. Because one agent can act for many users, a site cannot tell whose preferences, credentials or purchase authority sit behind a given request. The post's illustrative assistant, Ezra, runs errands for four clients; the grocer loses the personalization he once had, the electronics store cannot prove which client authorized a phone purchase in a friendly-fraud dispute, and a bank facing a password-guessing run can only ban the shared assistant, taking four innocent clients with it. The post lists concrete consequences: age and geography verification that agents cannot satisfy, dispute evidence that no longer exists because purchases originate from the agent's infrastructure rather than the user's device and address, unreliable agents generating refunds and support cost, constant polling of stock, price, search and support endpoints, free trials and promotions consumed, uncertain cardholder authorization, and content read or extracted without a visit, ad impression or purchase. In nearly every case the site absorbs the cost.
Cloudflare says its aim is an internet where all users have equal or greater site access than before agents. It accepts that agents should be able to act without a human present at every step, and even to browse privately, but argues that some actions — logging into a bank is the example given — should require additional verification.
It dismisses two popular remedies. Passing a CAPTCHA to the user proves only that some human is present at that moment; it does not separate one user from another and says nothing about the rest of the agent's activity, and operators are incentivized to strip that friction anyway. Tightly scoped access controls fail because agents are built to do anything their user could, so scopes collapse to "everything," and even a narrow scope records permission rather than assigning responsibility when things go wrong.
On identities, Cloudflare favors Web Bot Auth (WBA), in which an agent signs its requests, identifying itself and implicitly its operator. It notes a gap for chained intermediaries — Agent A using Agent B inside Browser C — which must be visible to prevent policy circumvention. It describes identity as a spectrum with data minimization: nothing at all; a session identifier, such as the one Poppy uses, that lets a site build reputation within a session; an account-level identifier stable across sessions, enabling binding a site account to an agent account; and, for high-risk or compliance-heavy actions, a unique per-user identifier from a third party that persists across operators, accounts, agents and sessions. Government IDs, third-party verification and decentralized approaches such as Proof of Human are cited as sources for that last tier. Privacy Pass and MOLE are put forward as ways to establish trust without identifying anyone; MOLE is described as the intended successor to Private Access Tokens and aims at rate limiting without user-specific identifiers, in the same spirit as the payment-focused x402 protocol. OAuth is cited for conveying account identity.
Separately, some actions require a claim about the real person — being over 18 or being the cardholder — which an operator or third party may already have verified at signup. Cloudflare argues sites need a way to request such claims and have them supplied. It also wants a machine-readable way for sites to publish which agent actions are permitted, since agents currently learn the rules by probing and reading block responses, wasting resources on both sides and sometimes making honest agents look hostile. Robots.txt covers where crawlers may go but not what agents may do. The IETF's AI Preferences Working Group is working on preference expression, but its scope is use preferences such as training, inference and search, and Cloudflare says no group yet covers the broader need.
Locally hosted or on-device agents, such as OpenClaw, are treated as a different case: they look like plain automation and have no single party to build reputation with or hold accountable. Asking them for a unique ID is one option, but Cloudflare says this is not TAD's initial focus.
For customers who do not want to wait on standards, Cloudflare says three things are available as of this week. Security Analytics now reports Personal Assistants as their own traffic category, and a new widget lists every agent visiting a site with request counts, paths hit and errors encountered. Enterprise Bot Management customers can use Business Insights to compare how much each operator crawls against the traffic it sends back. A third capability, setting access per agent from the dashboard, appears to be in the same rollout.
The practical takeaway for site operators is that agent traffic needs to be measured and classified before it can be priced or policed, and Cloudflare is positioning its analytics as that first step while the identity layer is still being standardized. For teams building agents or runtimes, the near-term implication is that signed requests and explicit identity tiers are likely to become table stakes for access to higher-value actions; that is my inference from the direction of the proposals rather than a stated requirement. A likely trade-off is that the richer identity tiers buy accountability at the cost of the privacy and frictionlessness that make agents attractive in the first place, which is why the post keeps emphasizing data minimization and privacy-preserving reputation mechanisms. The coalition's success will hinge on whether operators, identity providers and payment networks converge on the same signals rather than shipping vendor-specific variants.
Why it matters: Developers building agents or agent runtimes may soon need to sign requests and expose operator identity to reach ordinary sites, and product teams running sites will need per-agent analytics and policies to separate legitimate assistants from abuse. Cloudflare's per-agent traffic breakdowns and Business Insights give operators immediate tooling, while the standards work determines whether those controls become portable across vendors. Because the coalition includes payment providers and identity companies, the identity and claims question is likely to be settled first in commerce flows, in my reading.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
AI agents act for millions, but servers see intermediaries that strip user visibility. This post defines the problem space and introduces a collaborative approach to safely identifying, allowing, and managing this automated traffic.