On August 4, Cloudflare gave AI agents a wallet and a name. The announcement covered two things: Cloudflare Wallets, a programmable wallet for agents, and cloudflare.pay, a handle that tells a merchant which account an agent is acting for.
Cloudflare proposed Web Bot Auth in May 2025, shipped signed agents in August 2025, and by October 2025 had Visa, Mastercard, and American Express building on it. This has been in production for 12 months while most sales teams weren’t looking.
If you run outbound, this is closer to your desk than it looks. Three parts of a modern sales stack are automated traffic that somebody else’s system now evaluates: the domains and mailboxes you send from, whatever tooling collects and verifies your prospect data, and the LinkedIn automation sitting between the two.
The web is building machinery to tell wanted automation from unwanted automation, and outbound is automation.
What’s genuinely new is narrower than the headlines suggest, and it’s still mostly a promise. Here’s what shipped, what didn’t, and the part that lands on your desk.
What landed during Agents Week
Cloudflare ran Agents Week from Monday, August 3 through Friday, August 7, 2026. Wallets came out on Tuesday.
August 4: Wallets, handles, and what you can actually use today
Handle reservation opened on August 4. Funding an Account Wallet, issuing Virtual Wallets to individual agents, and letting those agents spend are all described as arriving in the coming months.
A human owns an Account Wallet that holds stablecoin funds. That owner delegates capped spending to per-agent Virtual Wallets operated through API keys, with a weekly allowance, an approved merchant list, and a ceiling on any single transaction. Handles look like research.example.cloudflare.pay.
The rest of the week: Access models, attribution, and continuous trust
August 5 brought the Agent Access Model, a framework for agents accessing resources on a user’s behalf, plus identity-aware analytics that attribute AI activity back to specific users and systems.
Friday August 7th is the one worth your attention. In Unveiling good and bad behaviors on the Agentic Internet, Cloudflare reframed bot mitigation around continuous trust instead of one-time risk scoring.
A visitor no longer gets classified once at the door. It gets evaluated on behavior, repeatedly, for as long as it keeps making requests.
Sit with that for a second if you run any kind of automated prospecting. A data vendor that scraped clean for two years can be reclassified next month based on how it behaves, without warning and without anyone telling you. The same is true of any browser automation your team runs.
Passing the first check stopped being the finish line, and the thing being evaluated is a stream of behavior rather than a one-time credential.
The correction most coverage got wrong
Identity shipped in May 2025
Web Bot Auth lets an automated client sign every HTTP request with a cryptographic key, using HTTP Message Signatures under RFC 9421, an Ed25519 key per agent, a Signature-Agent header, and a published directory of public keys. Cloudflare proposed it in May 2025.
In August 2025 it announced the first cohort of signed agents, with Block’s Goose, Browserbase, and Anchor Browser among the named partners.
Adoption followed fast. AWS WAF Bot Control added support in November 2025. Vercel, Akamai, Stytch, and Shopify followed. Google documents an experimental implementation for its own agent traffic.
The payment networks got there in 2025 too
Here’s the detail almost nobody repeated in August. On October 14, 2025, Cloudflare announced it was working with the major payment networks on an authentication layer for agentic commerce, all of it built on Web Bot Auth.
Visa co-developed the Trusted Agent Protocol and folded it into Visa Intelligent Commerce. Mastercard built it into Agent Pay. American Express committed to using it in its agentic commerce program.
Which means the “can this agent prove who it is” question was settled infrastructure before the wallet existed. August added the money, on paper.
How much of the web is already automated
57.5% of HTML page requests, June 3, 2026
On June 3, 2026, Cloudflare CEO Matthew Prince posted Radar data showing automated systems generating 57.5% of HTTP requests to HTML content worldwide, against 42.5% from humans.
He’d predicted the crossover at SXSW in March and expected it around the end of 2027. It came roughly eighteen months early.
Use the full sentence when you quote this internally. The measurement is HTML page requests on Cloudflare’s network, not all internet traffic. NBC News reported 57.4% from the same dataset, and Prince himself called the data “a bit messy” in replies to his own post.
The number that keeps you honest
Over the same period, Cloudflare’s all-traffic bot metric sat near 35.2%. Both figures are accurate because they count different things. HTML requests exclude images, scripts, API calls, video, and gaming, all of which humans generate at much higher rates.
Anyone who tells you “bots are 57% of the internet” has dropped the denominator and inflated the claim by a factor of roughly 1.6. Which matters when you’re the one presenting the slide.
Shipping ahead of the standard
A working group exists. It hasn’t adopted anything.
Call Web Bot Auth a standard in a room with an engineer and you’ll get corrected. The IETF chartered a working group for it in early 2026, following a Birds of a Feather session at IETF 123, co-chaired by David Schinazi and Rifaat Shekh-Yusef. Milestones were set for a standards-track specification and a best current practice document on key management.
As of August 12, 2026, that working group had adopted zero documents. Everything in flight is still an individual submission. Meanwhile Cloudflare, AWS, Akamai, HUMAN, and Vercel verify these signatures in production every day, deciding which agents reach which websites.
What that gap costs, concretely
An example rather than a warning. The architecture draft revised on August 6, 2026 moved the work to Standards Track and states that signers must send the Signature-Agent header in dictionary form.
Cloudflare’s published verification rules still instruct implementers to use a structured string and treat the dictionary form as a reason to fail the request.
Implement the current draft and Cloudflare will reject the request. Implement to Cloudflare’s documentation and you’re coding against a spec that has already moved past you.
Either way somebody on your team loses a morning to it. That’s what “shipping ahead of the standard” costs in practice, and it’s the reason this belongs on a watch list for now.
What this means from an outbound seat
Everything above is infrastructure news. Here is the translation into work your team either does or does not do this quarter.
The three surfaces your outbound stack exposes
Outbound touches the agentic web at three points, and most teams only think about the first one.
- Sending infrastructure. Your domains, your mailboxes, your authentication records. This is the surface receivers judge before a human reads a word.
- Automated collection. Enrichment vendors, verification services, browser automation, anything that fetches a page or an API on your behalf to build or clean a list. This traffic is now subject to continuous trust evaluation at the CDN layer.
- Social and dialer automation. LinkedIn tooling and AI calling systems sit under platform-specific rules rather than Web Bot Auth, though the direction of travel is identical: platforms want to know which automation is acting for whom, and they are getting better at telling.
A team can have flawless email authentication and still be exposed through a data vendor nobody has audited in eighteen months.
Your sender identity, as a machine sees it
Run this audit on your own domain. Four checks, none of which need an engineer for more than an hour.
- Does your primary domain have SPF, DKIM, and DMARC passing and aligned? A machine evaluating your outreach starts where the mail server starts.
- Is your public web content readable without JavaScript execution? Agents that can’t render your page can’t verify your claims about your company.
- Do your pricing, integrations, and security pages state facts in text rather than in images or PDFs?
- If your team runs any automated browsing or scraping, do you know whether those requests are signed? If the answer is “the vendor handles it,” get the vendor’s answer in writing.
Point four is the one people skip. It’s also the one where the honest response from most vendors right now is “we don’t sign, and nobody has asked us before.”
Enrichment gets a paper trail
Data vendors have historically operated on unsigned traffic and inference. With verification moving into CDNs and WAFs by default, unsigned automated collection becomes visible, attributable, and easy to rate-limit.
Ask your enrichment providers two questions this quarter. Do they sign their requests, and under whose identity? Vendors with a good answer will give it immediately. The ones that get vague are telling you something.
There’s a commercial consequence worth naming. If a provider’s collection gets rate-limited or blocked at scale, your list quality degrades quietly. Bounce rates climb, complaint rates follow, and the first symptom you notice is a deliverability problem that has nothing to do with your sending.
Buyer-side screening, stated as the guess it is
Now the directional part, and it’s labeled that way on purpose. There’s no published data on how often buyer-side agents currently screen inbound sales outreach. None. Anyone quoting you a percentage is making it up.
What we can point at is where the infrastructure leans. Agents that hold funds and identify themselves to merchants are agents that can also filter, rank, and decline. Cloudflare’s own framing for agents that stay anonymous is worth borrowing: an unidentified caller has to work harder to prove itself. Apply that to a seller emailing a buyer whose assistant reads the inbox first.
Where AnyBiz fits in this
We build an AI sales agent that runs outbound end to end: finding prospects, reaching them across email, LinkedIn, and phone, handling replies, and booking meetings. Which makes us one of the automated systems this article is about, and gives us a direct stake in how the agentic web decides what to trust.
FAQ
Did AI agent identity launch in August 2026?
No. Cloudflare proposed Web Bot Auth in May 2025 and announced its first signed agents that August. The August 2026 announcement covered payments and handles.
Can AI agents actually pay for things now?
Not yet, at least not through Cloudflare Wallets. Handle reservation opened on August 4, 2026, while funding, Virtual Wallets, and programmable spending were described as arriving in the coming months. If you’re evaluating this for a real workflow, build behind an abstraction and check status before you commit. Other players, including Visa and Mastercard, have their own agentic payment programs running on separate timelines.
Is Web Bot Auth an official standard?
It isn’t. An IETF working group was chartered in early 2026, but as of August 12, 2026 it had adopted no documents, so everything remains an individual submission. Production deployment across major CDNs is running well ahead of that process.
Are bots really the majority of web traffic?
Depends entirely on what you count. Bots reached 57.5% of HTTP requests to HTML content in June 2026. Across all traffic types, Cloudflare’s figure was closer to 35%.
Does any of this affect cold email deliverability today?
Indirectly, and through your data rather than your sending. Web Bot Auth governs HTTP traffic, not SMTP, so it has no direct bearing on whether Gmail accepts your mail. The connection runs through list quality: if the vendors building your lists get rate-limited as unverified automation, your data goes stale, bounces climb, and deliverability follows them down.
What should a sales team do about any of this?
Fix your domain authentication, make your commercial pages machine-readable, and find out whether your data vendors sign their requests. Everything else is a watching brief for now.
