You are reviewing signup analytics one afternoon when a cohort stops making sense. These accounts never open the pricing page, register no scroll behavior on the landing page, and move from the API documentation to a generated key in under a minute, in the same order every time. You suspect fraud, and every check comes back clean, with valid payment methods and real companies behind the accounts. The same week, support receives its first ticket written by an assistant on behalf of a customer, a single precise question about whether the rate limit applies per key or per organization. Nothing is broken, and revenue looks normal. The change is quieter than any outage: a new kind of user has arrived, and it is not a person.
This chapter opens the part of The Frontier about products whose users are agents, and it starts with who these new users are, why they showed up now, and what they change about your job.
The three waves of agent traffic
Agent traffic is not one phenomenon, and treating it as one produces bad calls, because a scraper you want to block and a paying customer's assistant can look almost identical in a log. The traffic arrives in three waves, and each wave touches a different part of your product.
- Crawlers that read. Answer engines and research assistants fetch your documentation and help pages, then answer the person's question directly, so the visit you used to earn never happens. Infrastructure providers that measure this traffic report automated reads climbing steeply relative to the human visits referred back, which means your content works harder every month for readers who never see your navigation, your ads, or your brand.
- Shoppers that choose. An assistant is asked to solve a problem, finds a handful of vendors that fit, compares them on capability, limits, and price, and completes a signup or a purchase on the person's behalf. It reads structured facts rather than positioning, carries no habit or brand memory from last time, and runs the same comparison again at the next renewal. The strange cohort in the opening scene is this wave landing.
- Operators that act. Agents finish jobs inside your product, either by clicking through the interface a person would use or by calling your tools directly, and they act on authority a customer delegated to them. This wave carries the most consequence, because an operator is not reading about your product or selecting it; it is working inside it with a real customer's permissions.
Agent traffic arrives in three waves, crawlers that read, shoppers that choose, and operators that act, and each wave is bigger and more consequential than the one before.
The waves overlap rather than queue. For most products the crawlers are already here, the shoppers are arriving wherever a purchase can complete without a phone call, and the operators are early but no longer hypothetical, as the support ticket in the opening scene shows.
Why this is the mobile moment again
A decade and a half ago, products won by designing for a screen their own teams barely used. The teams that treated the phone as a first-class user, rethinking flows and stripping pages down rather than shrinking the desktop site, took markets from incumbents who treated it as an inconvenient viewport. The parallel discipline now is designing for a user your team will never be: one that reads structure instead of layout, compares without loyalty, and never sees the brand you spent years building. Nobody on your team experiences your product the way an agent does, and the mobile winners already proved that a team can design well for a user nothing like itself.
The infrastructure just stabilized: why agents arrived now
Traffic needs rails to travel on, and the reason the waves are arriving now is that the connective infrastructure settled within roughly the last two years.
- A common way to connect tools to models. The Model Context Protocol, a standard that lets any model client call any product's tools, moved from one vendor's open-source project to shared infrastructure governed by a neutral foundation, with adoption measured in hundreds of millions of downloads. When every major client speaks the same tool language, exposing your product to agents stops being a per-vendor integration.
- A common way for agents to work together. An agent-to-agent protocol, which lets one agent discover another and hand it work across company boundaries, reached its first stable version this year under the same style of neutral governance. It extends across organizations the coordination we covered inside one product in Why one agent stops being enough.
- A common way for agents to pay. Payment networks and major commerce platforms shipped protocols that let an agent complete a purchase carrying a mandate, meaning a verifiable record of what its principal authorized it to buy and spend. That closes the last gap between an agent that can pick a vendor and an agent that can become a customer.
The pattern matters more than any version number: the standards moved to neutral ground, the large platforms committed to them, and each protocol removed a reason an agent could not be a reader, a buyer, or an operator of your product.
What agent users change about your job
Your next users will read your tool descriptions, not your homepage: they compare without loyalty, act on delegated authority, and never see the interface you polished.
That reversal lands on work you already own.
- The tool interface becomes a storefront. When an agent checks whether your product can do a job, what it reads is your tool names, descriptions, and parameters, so definitions you treated as internal hygiene start doing the persuasion your homepage used to do. Every tool you expose is also authority you hand a guest, the ground we covered in The action surface: every tool is delegated authority.
- Docs become the sales floor. Crawlers and shoppers assemble their entire picture of your product from documentation, so its accuracy, structure, and machine readability move from a support cost to a revenue channel.
- Delegation becomes the consent flow. When an operator acts with a customer's permissions, whose keys it holds and what it may do stop being implementation details and become customer-facing product design, work that builds on Identity: whose keys your AI holds.
- Pricing meets a buyer that never skips the comparison. A shopper re-runs the vendor evaluation at every renewal, so pricing that survived on customer inertia gets re-examined on every cycle.
The chapters ahead take these in turn: making your product findable by machine readers, designing delegation and consent, setting guardrails for guests acting on borrowed authority, handling agents that arrive through the human interface, and working out what the business looks like when a bot is the buyer.
Try it now
This drill takes about half an hour, spends nothing, and tells you which wave has already reached you.
Get your specimen. Choose a product you know well, ideally your own, and open its public documentation in a private browser session with no cookies and no login.
Read like a machine. Ignore every marketing page and work only from the docs and API references. Write down whether a machine reader could establish, without rendering a single page of marketing, what the product does, which actions exist (endpoints, tools, or operations), and what they cost. Note wherever the answer lives only in a screenshot, a video, or an interactive pricing widget, because those spots are invisible to the second wave.
Check for fingerprints. Open your traffic analytics or server logs for the last month and look for agent fingerprints: user-agent strings from known AI clients and crawlers, sessions with no scroll or pointer events, and paths that run straight from documentation to API key with nothing in between.
Classify the wave. Write one sentence stating which of the three waves has already arrived for this product, with the evidence beside it, and keep the sentence, because the coming chapters each start from where your traffic actually is.
Chapter Summary
- Agent traffic arrives in three waves: crawlers that read your docs, shoppers that choose among vendors, and operators that act inside your product, each bigger and more consequential than the last.
- Crawlers answer the user's question directly from your content, so the visit you used to earn often never happens.
- Shoppers compare structured facts without brand loyalty and re-run the comparison at every renewal.
- Operators act on authority a customer delegated, which makes identity and permissions customer-facing product design.
- The rails settled recently: a shared tools-to-models protocol under neutral governance, an agent-to-agent protocol at its first stable version, and payment protocols that carry a purchase mandate.
- This is the mobile moment again: design for a user your team will never be, before your competitors do.
- The tool interface becomes a storefront, docs become the sales floor, delegation becomes the consent flow, and pricing meets a buyer that never skips the comparison.
- Do not let scraper defenses turn away delegated agents; a blocked assistant reads to the customer as a product that failed.
The next chapter, The tool interface is the storefront: design actions agents choose, starts where every agent starts reading: your tool definitions.
Sources
- Cloudflare (2025). Radar and blog reporting on AI crawler and bot traffic, including crawl-to-referral ratios.
- Anthropic (2025). Model Context Protocol documentation and governance announcements, with industry coverage of the protocol's donation to a neutral foundation (last verified July 2026).
- Linux Foundation (2025). Announcement of the Agent2Agent protocol project; Agent2Agent protocol stable release documentation (last verified July 2026).
- Google (2025). Agent Payments Protocol announcement.
- OpenAI and Stripe (2025). Agentic Commerce Protocol announcement.