In March, Cloudflare’s CEO advised an viewers at SXSW that bot visitors would surpass human visitors on-line by 2027. On June 3, he posted that it had already occurred: automated techniques now generate 57.5% of HTTP requests to net content material, towards 42.5% from people, per Cloudflare Radar, the primary machine majority the corporate has recorded. The motive force is agentic AI: a single agent performing a buying job can go to a thousand instances extra websites than an individual would. These numbers describe the online I’m truly transport into: we construct measurement and attribution infrastructure for the agent financial system, working instantly with the retailers and companies attempting to determine make their companies legible to LLMs and brokers. From that vantage level, virtually each assumption the web was architected round is quietly breaking.
The assumptions which might be breaking
Price limits, classes, identification, and analytics have been all constructed round a single assumption: a human at a browser, leaving a constant footprint as they moved by way of a web site. That assumption now not holds, and it doesn’t fail in a single course. It fails in lots of.
Once we examined how each main AI system truly interacts with web sites, we discovered not less than six distinct architectures in manufacturing: uncooked HTTP fetches with no JavaScript or analytics, headless browsers operating from datacenter IPs, browser extensions like Claude in Chrome operating contained in the person’s personal session with their cookies and auth state, screenshot-based brokers like ChatGPT Agent Mode working distant Chrome on Azure VMs, OS-level pc use that controls purposes past the browser, and rising protocols like WebMCP that bypass the web page interface altogether. Some are seen in server logs however not analytics. Some are seen in analytics however indistinguishable from people. Some are invisible in all places. The identical purchaser may use three of them in a single buy, producing three totally different footprints throughout one journey. Identification is equally fragmented: an agent in a browser extension carries the person’s actual cookies and auth, whereas an agent on a distant VM carries none of it.
What truly works in manufacturing
Earlier than OpenLens, I used to be on the Elementary AI Analysis crew at AWS AI Labs, and earlier than {that a} analysis technician at Caltech. From that vantage level, the factor that strikes me most concerning the present public dialog on agent reliability is that it treats brokers as one class. They aren’t. Brokers that contact clients and brokers that don’t are fully totally different issues with fully totally different playbooks. Conflating them results in dangerous selections on each.
Inside brokers, those your engineers and operators use to ship sooner, draft work, parse logs, summarize conferences, are pure productiveness. The output will get reviewed by a human who owns the consequence. Probably the most counterintuitive factor I’d inform engineering leaders proper now’s to level AI at your hardest, most important work, not your best. Most individuals underestimate what present fashions can already do, and silo them off on duties they don’t truly must confirm. That’s precisely backwards. The bottleneck is now not mannequin functionality. It’s human self-discipline about verifying output earlier than it ships. Groups that delegate trivial duties and confirm nothing produce slop. Groups that take swings at laborious issues and rigorously examine the output produce leverage.
Buyer-facing brokers are a special class solely. The human possession chain breaks the second the agent is the interface. That’s its personal engineering self-discipline with its personal playbook, and many of the public dialog muddles the 2. The leaders price listening to are exact about which downside they’re truly fixing.
Constructing on rented land
Constructing on prime of a frontier mannequin is a bit like constructing on prime of another person’s rented land whereas they renovate the home. Fashions replace each few weeks. Pricing strikes. Generally particular conduct on a selected immediate adjustments with out warning.
Two examples from our personal stack. We use Haiku to parse LLM responses for model mentions. The immediate is written to explain what we wish clearly, to not exploit quirks of a selected mannequin model. As Haiku improves, the parser improves with it. We should not have to the touch the code.
Final yr, on an earlier model of our stack, we laborious coded the API identify for Claude 3.5 Sonnet in roughly ten totally different locations. When that mannequin was deprecated mid-deployment, the failures cascaded by way of a pipeline that touched customers. We rebuilt the abstraction layer afterwards.
The lesson is to write down to functionality, to not variations. Code that is determined by what a selected mannequin does as we speak breaks when that mannequin adjustments. Code that describes what you truly need will get higher because the fashions do.
What I feel this all factors to
The subsequent technology of merchandise, shopper and B2B, will likely be judged on how gracefully they deal with agent visitors, not simply human visitors. Not in some summary “AI-ready” sense however concretely: can brokers uncover your product, authenticate with bounded scope, carry out the motion the person truly delegated, depart a clear audit path, and do all of that with out degrading the expertise for the people who nonetheless present up. Most corporations I discuss to haven’t began occupied with this, however the ones which have are going to compound a bonus that could be very laborious to make amends for later.
