Agents as intermediaries

An AI Agent acting on the web may look to a Site like an ordinary human completing a task, or like a single visitor fronting for many Users. Either way, the Site can no longer recognize, serve, or hold accountable the people behind the Agent, while the Site absorbs the cost when things go wrong. The argument here is that expanding Agent access to Sites depends on new mechanisms that preserve the Site–User relationship: Identity, Reputation, and Accountability, or, in plain terms, whether the actor is truthful about who they are, whether that actor is trustworthy, and who answers for the action.

"Personal AI Agents" are automations that take instructions from a human User and pursue a goal by controlling a cloud computer or browser, usually hosted and maintained by an Operator as a managed service or consumer offering. Examples include Muse from Meta and dots from OpenAI. An Agent can take any online action its User could, minus the limits that constrain humans: it does not tire, it runs continuously, it acts in parallel, and it tolerates friction no person would. Sites plan capacity and make money on the assumption that those limits exist.

Remotely hosted Agents are, fundamentally, intermediaries: anything that sits between a User and a Site so the Site sees it instead of the User. Cloud browsers and Proxies, including VPNs, are intermediaries as well; what has changed is the trajectory. Agentic traffic is growing fast and may eclipse traditional human browsing in the near future. Blocking intermediary traffic and telling Users to disable the service stops working when most traffic is agentic. Once blocking is off the table, the Site faces the age-old intermediary problem: it loses direct information about the User and must decide whether to trust the intermediary without it.

Agents hosted somewhere other than the User's device are the primary subject here. Local agents (e.g. OpenClaw) and Agents running in the User's own browser on their own device fall outside scope (see "On open source and local Agents" below).

image5.png
Without Agents, the User interacts with the Site directly.
image6.png
With Agents, the Site sees the Agent, not the User, and one Agent acts for many Users.
image7.png
Sometimes several intermediaries sit between the User and the Site.

The post rests on one principle:

The Intermediation Principle: The business relationship between Users and Sites, and each participant's security responsibilities, must be upheld regardless of intermediation.

Sites need tools to recognize, serve, and hold accountable a User acting through an Agent exactly as if that User had visited directly. The Agent and its Operator must be recognizable and accountable too, since an Agent can do what no User could: Agents scale human access almost without limit, creating new risks even for previously benign behavior.

Ezra, the personal assistant

Consider Ezra, who runs errands for people in her city: her best client is Amy, a local CEO, and she also works for Bob, Charlie, and Debra.

The corner grocer used to know Amy's preferences and set aside what she'd like. Now only Ezra comes in, carrying orders for four people, and the grocer cannot tell whose preferences are whose — yet he won't turn Ezra away, because she brings a lot of business. When Bob asks Ezra to buy an iPhone and later calls the store claiming he never authorized the charge and keeps the phone, the store must prove Bob asked for it, but the only person it ever saw was Ezra. When Charlie hands Ezra a list of possible passwords to try at his bank, the bank can see an attack in progress but can only ban Ezra — which also bans Amy, Bob, and Debra.

None of these owners has a problem with Ezra herself. The problem is seeing only Ezra, not the person she acts for. Replace Ezra with an AI Agent and the businesses with websites, and you have the web today. A Site owner wants enough visibility both to prevent bad behavior and to monetize the Site, and much of that visibility disappears with Agents.

ChallengeWebsite owner’s goal
Account sharingSites need to ensure that one paid subscription is not being abused to serve many Users. The Site sees one account and one client when accessed via an Agent.
Compliance and illegal activity

A Site must meet legal requirements to verify information about its Users. For example, some Sites have age or geographic restrictions.

Agents do not provide guarantees that can meet the bar for Sites to remain compliant, if they allow Agent access.

Friendly fraud

Friendly fraud is when the User approves an Agent's purchase, then disputes it as unauthorized. Sites may win with circumstantial evidence: the same device, IP, and address as the User's past undisputed purchases. Agent purchases come from the Agent's infrastructure, so that evidence doesn't exist. Even when the Site wins, the dispute itself costs money.

Sites need to maintain records of checkout behavior for fraud prevention purposes. Maintaining these records helps Sites reduce overall friction from additional verification, and results in a lower transaction decline rate, and overall greater number of checkouts.

Human verification is irrelevant

A Site wants to be able to distinguish the type of Users it has. CAPTCHAs and other human verification check whether the entity controlling the browser is a human or not.

Agents are never expected to be humans. Solutions that pass this signal through to the human User add no new information and just make Agents less efficient.

Identity spoofingA Site wants to limit access to content to human Users or impose restrictions on what particular Users can access. Agents often run on residential IPs and use real Browsers, are mixed with human-directed traffic, and take efforts to not distinguish themselves from human traffic. Attackers copy a well-known Agent's user agent string to get the same allowances.
Identity verification breaksA Site may require additional verification of the User’s identity, or other information, in order to provide access to sensitive operations or content. Liveness checks, KYC ID checks, and similar assume the person present is the person being verified. The Agent is never that human. The Site has to accept verification passed through by the Agent (which weakens the check), reject Agents, or drop the check.
Misalignment

Sites may want to protect themselves from erroneous actions taken by newer or more experimental Agents.

Agents make mistakes: the wrong item, the wrong date, duplicate orders. Sites may absorb the cost through refunds and support to keep customer goodwill, with no way to tell which Agents are reliable.

Policy circumventionThe Site is looking to impose a policy (e.g. block, allow only partners, rate limit) that is based on identity, but Agents don't always identify themselves, and they often don't bring information about the User they act for.
Resource exhaustion

Sites that provide free services or that rely on User engagement were often designed with an assumed limit on User behavior.

Agents poll constantly (e.g. stock and price checks), hit expensive endpoints (e.g. search, inventory, or support chat), and use up free trials and promotions.

Risk-based authentication breaksA Site wants to maintain its risk-based authentication system, which expects the device, location, and behavior to match the User. Agents run on cloud infrastructure and look unusual every time.
Traffic comminglingSites would like to protect themselves and User accounts from abuse or fraudulent activity. When one User abuses a shared Agent for credential stuffing, scraping, or card testing, the Site can detect the abuse but can only block the Agent, which also blocks all of its legitimate Users.
Unauthorized card users

Sites and payment providers need to protect cardholders from unauthorized transactions, while minimizing dispute costs and other costs related to fraud.

When an authorized User makes a purchase on a card, the cardholder is responsible. If the cardholder hands the card to someone who isn't an authorized user, they're still responsible for having handed it over. Agents have no authorized-user status, so the Site can't tell whether the User stands behind this purchase. In practice, when the User disputes it, the Site pays.

Uncompensated consumption

Sites that monetize content require visibility and control over who accesses that content in order to apply the correct monetization model.

The Agent reads or extracts content (articles, product data, reviews) to answer the User somewhere else. The Site pays to serve it and gets no visit, ad impression, or purchase.

User attributionSites offer additional capabilities to consenting Users regarding personalization, recommendations, and loyalty programs, using analytics tied to the User's session and device. Agent traffic either groups many Users into one client or shows up without the User's session at all.

In nearly every challenge above, the Site bears the primary cost — an outcome that follows from how trust works when an intermediary is involved.

The goal: equal or better access

The aim is an Internet where all Users have equal or greater access to websites than before. Pre-Agent, Sites mitigated these risks with techniques including fingerprinting the User's device, session, and history. Users could still browse in private browsing mode, and Agents should be able to do that as well — though they probably won't log in to a bank without additional verification. None of this requires an Agent to act like a human or a human to be present for every action. It requires that Sites have the means to obtain or require information about the User in order to hold them accountable where applicable, with or without an Agent.

Why prior proposals fall short

Passing a Site's CAPTCHA through to the User proves only that some human is present at that moment ("human in the loop", or HITL), and nothing more. HITL primarily prevents mistakes from misalignment — an issue between Operator and User — and resolving it does not resolve most of the problems above. It cannot tell Bob from anyone else (Traffic commingling), and it says nothing about whom the Agent acts for the rest of the time. Since Agents exist to remove friction and HITL adds it back, Operators are incentivized to remove or bypass it for the User.

Tightly scoped access controls fail for a similar reason. Agents are built to do anything a User could, so in practice they request and receive permission to do everything — a scope of "everything" tells the Site nothing. Even a narrow scope says only what the Agent was allowed to do, not who answers when the Agent gets it wrong, which is the question the Site actually needs answered.

Standards work already in progress

The gap between what Sites need and what Agents can currently do is not going to be closed by vendor-specific mechanisms. It needs open, interoperable standards grounded in operational experience, and a fair amount of that work is already underway. Many of the problems described can be addressed by applying existing standards, and they fall into a few recognizable buckets.

Protocols for access without identifiers

Agents and Sites can enter into explicit protocols rather than leaving the Site to infer intent from behavior. MOLE is one upcoming approach aimed at rate limiting without User-specific identifiers, letting Sites move away from implicit identities toward explicit protocols that accomplish the same goal. The x402 protocol follows a similar philosophy for payments: enabling payment flows on the web while remaining privacy-preserving. Where existing protocols fall short, new ones may need to be designed.

Agent identity via Web Bot Auth

For Agent Identity, Web Bot Auth (WBA) is the preferred method. It lets an Agent sign its requests, giving it a secure way to identify itself and, implicitly, its Operator. One open problem: when Agent A uses Agent B inside Browser C, the Site needs to know that chain of intermediaries so it can correctly understand and control access by Agent A — otherwise Policy circumvention becomes possible. Session, account, and User-specific identifiers may also be necessary depending on context, and work with the relevant standards bodies is exploring how to convey them.

The identity spectrum

Identity is best understood as a spectrum trading different degrees of information about a user, ideally in line with data minimization. It answers “who are you?” in service of the follow-up questions “are you who you say you are?” and “can we trust you?”. In many cases the first question doesn't need answering at all: Privacy Pass and MOLE can establish trust without identifying the user. Private Access Tokens was an early attempt at trust without identification, and MOLE is intended as its successor. Where graduated disclosure is needed, the forms generally look like this:

  • Nothing — the user acts with no identifiers at all.
  • Session — the Site can link requests within one session to a single agent and build reputation from it. Poppy, for example, defines a session ID for this purpose.
  • Account — requests are attributed to the same User across sessions, e.g. a stable identifier assigned by an Operator or third party that persists over time, enabling things like binding a Site account to an Agent account.
  • Unique — a per-User identifier provisioned by a third party that persists across Operators, accounts, Agents and sessions. This suits high-risk actions, limited access to expensive resources, or strict compliance obligations; government IDs, third-party verification and Proof of Human are examples of technologies that can supply it.

What a Site actually needs to answer is “is the entity behind this request allowed to do this action at this time?” Several pieces already map onto that question: WBA carries an Agent's platform identity, MOLE carries User reputation, OAuth carries account identities.

Claims and site policies

Some actions need a fact about the real person — being over 18, being the cardholder (see Identity verification breaks and Friendly fraud). The Operator often verified this at sign-up, or a third party already holds it, so Sites need a mechanism to request such claims and Operators or third parties a way to supply them.

Sites also need a machine-readable way to publish which actions Agents may perform. Today an Agent discovers a Site's rules by trying something and reading the response, usually whether it got blocked — wasteful on both sides and sometimes making honest Agents look malicious. Robots.txt covers where a crawler may go but is not suited to expressing what Agents may do. The IETF's AI Preferences Working Group is developing preference-expression mechanisms, but its focus is on use (training, inference, search), and no group yet addresses the broader need for the preferences discussed here.

Local and open source Agents

Agents not run as an Intermediary — on-device Agents, or locally hosted ones such as OpenClaw — raise different concerns. They look like automation, and they are, with no single party to build reputation with or hold accountable. Requiring these Agents to provide a Unique ID is one option, but it is not the initial focus of Trusted Agent Delegation.

Trusted Agent Delegation and next steps

The plan is to bridge running code and emerging specifications across work like Web Bot Auth, AIPREFs, OAuth, WIMSE and MOLE, partly by co-authoring and refining open protocols such as WBA and MOLE and deploying implementations of them at scale. Collaboration is ongoing with major Operators, identity verification companies, payment providers, merchants, publishers, proxy providers and cloud browsing companies, aligning operational requirements directly with active IETF work.

To give that effort a home, the Trusted Agent Delegation (TAD) coalition is being announced. TAD gathers representatives from across the ecosystem to discuss the problem set, examine existing protocols, and bring better-defined problems to the IETF as needed, with side meetings at the IETF and workshops to follow. Builders of agents and agent runtimes, edge and cloud platform operators, and people managing commercial or compliance infrastructure are all invited to participate.

What is available now

The open, interoperable foundation starts with TAD, but Cloudflare customers don't have to wait for new protocols to regain visibility and control. Three concrete steps are available now:

  • See which Agents visit and what they do. Security Analytics now treats Personal Assistants as a separate traffic category, and a new widget lists each Agent on the site, its request count, the paths it hits and the errors it encounters. Enterprise Bot Management customers can use Business Insights for the wider picture of AI crawlers, including how much each Operator crawls versus the traffic it returns.
  • Set per-Agent access. AI Crawler and Agent access policies let you allow an Agent site-wide, restrict it to pages without ads, or block it outright, independently of your Search and Training policies.
  • Tie an Agent to its user and act on it. With Account Abuse Protection (early access), each login and sign-up shows whether a personal Agent made the request and which one. If an attacker abuses a shared Agent on one account, a WAF rule on its Hashed User ID mitigates that account without blocking the Agent for every other legitimate user — a concrete first answer to Traffic commingling.

These are initial building blocks, with more to come as standards mature. Companies with a more restrictive position on Agents are invited to share their problems directly with their Customer Support team or at [email protected]; the aim is an Internet where site owners have visibility and control, and Agents get the maximum reasonable access within those controls.