TollBit’s data tells a story that most people in tech haven’t fully absorbed yet. At the start of 2025, roughly 1 in every 200 website visits came from an AI bot. By late 2025, that number was 1 in 31, and human traffic declined quarter over quarter throughout the year. Aaron Levie put it plainly this week: AI agents are becoming “the primary users of the internet.” He didn’t put a date on the crossover, but the direction of travel is already clear.
The implications are not abstract. They are being built, right now, by founders who are asking a different kind of question: not “how do I get a user to click?” but “how do I build something an agent can discover, trust, reason over, pay for, and complete a workflow through?”
What’s Actually Being Built?
Last month, Meta quietly acquired Moltbook, a social network launched in January 2026 that was designed not for humans, but exclusively for AI agents. The platform described itself as “the front page of the agent internet.” By February it had amassed over 1.6 million active agents, with a ratio of 88 AI agents to every 1 human owner. No banner ads. No scroll-optimised feeds. No dark patterns. Just agents posting, commenting, and voting with no human in the loop. Meta moved fast enough to fold the founders directly into Meta Superintelligence Labs.
Moltbook is an extreme example, but it points to something real: a parallel web stack is forming. Cloudflare launched “Markdown for Agents” because the visually messy human web is expensive and inefficient for machine readers. Google introduced the A2A (Agent-to-Agent) open protocol because if agents are going to browse, compare, and buy, they need a shared language for transactions. Microsoft is building a publisher content marketplace for the agentic web. These are not side projects. They are early signs that one layer of the web is being rebuilt for human attention, and another for machine action.
Gartner has put numbers on the trajectory: traditional search engine volume will fall 25% by 2026 due to AI agents taking over query-answering, and organic search traffic to websites could be down more than 50% by 2028. UX pioneer Jakob Nielsen has called 2026 “the year of AI agents“, the moment the interface model shifts from Conversational UI (you ask, it answers) to Delegative UI (you want something done, the agent goes and does it).
Jakob paid $0.48 for this 5-minute AI job that resulted in two good infographics. If he hadn’t wanted to show you a selection, he could have paid less.
The Eyeballs Problem
For 25 years, the commercial web was designed around eyeballs. Pages were optimised for clicks, scroll depth, impressions, and ad inventory. But agents don’t care about branding flourishes, hero images, pop-ups, or clever copywriting. They care about clean structure, trusted identity, permissions, pricing logic, and reliable completion. They don’t “discover” products the way people do — they filter, rank, retrieve, and execute.
Agents don’t have eyeballs.
This is an existential challenge for the advertising model. When an AI agent books your flight, it doesn’t see the banner ads on the travel site. When it researches a product on your behalf, it skips the sponsored listings. When it reads an article to summarise for you, the display ad in the sidebar generates no impression. The entire scaffolding of digital advertising: CPMs, CTRs, retargeting pixels, view-through attribution, was built on the assumption that a human is emotionally available to be influenced.
The old web sold access to humans. The new web may sell structured access to machines.
That doesn’t mean advertising disappears — it mutates. The unit of value shifts from impression to outcome, from pageview to successful retrieval, from click-through rate to agent win rate. In an agentic web, the question is no longer only “how do I get the user to click?” but “how do I get the agent to choose me?” McKinsey has already sketched parts of that future: conversational marketplaces, commission-based models, payment-processing fees at the protocol level, and premium data access charged per retrieval rather than per view. In the human web, SEO became a discipline. In the agentic web, something closer to GEO (Generative Engine Optimisation), protocol optimisation, and machine trust design will matter just as much.
A new stack of standards is forming around this: llms.txt (a proposed convention for describing your site’s content to AI systems), structured product and content feeds, agent-friendly markdown, identity layers, permissioning, provenance signals, and payment rails. And the critical question of pricing — should agent traffic cost the same as human traffic? Per retrieval, per lead, per task completed, or per transaction initiated? is wide open. I wrote about AI Agents as the evolution of RPA, but now we’re moving to warp speed.
The Startup Landscape
You can already see companies forming around different layers of this stack.
Browserbase, Browser Use, and Notte are building browser infrastructure so that agents can actually operate on the existing human web — navigating interfaces that were never designed for machines.
Firecrawl (YC-backed) is building the web data extraction layer: search the web, navigate to any page, extract clean structured data, all from inside a fully managed browser environment purpose-built for agent workloads.
TollBit is building monetisation infrastructure for publishers and websites now facing significant machine traffic — paywalls and licensing models for AI access, so that content creators can charge agents rather than just block them.
Moltbook explored what social environments look like when the participants are agents, not people. Meta’s rapid acquisition of the company suggests that question has legs.
/dev/agents, founded by former Google and Stripe executives, is building a cloud-based operating system for AI agents — the infrastructure layer that lets agents collaborate, adapt, and function across environments. Their ambition is to be to AI agents what Android was to mobile.
Underpinning all of it: Anthropic’s Model Context Protocol (MCP), Google’s A2A protocol, and WebMCP are attempting to give the agent web a shared language — potentially the TCP/IP moment for autonomous AI systems.
The AI Agent ecosystem is growing rapidly, according to CB Insights:
The Challenges Are Real: Security, Compliance and Control
Building for agents is not the same as wrapping a chatbot around an old workflow. Agents need high-quality structured data, robust authentication, deterministic fallbacks, pricing logic, observability, and policy controls. They break easily in messy environments and the existing web is very messy.
They also introduce new attack surfaces: spoofed identities, prompt injection, data exfiltration, false actions, runaway loops, and compliance failures. If agents become economic actors with the ability to transact and commit on behalf of users, every weak interface becomes a vulnerability. The security and governance layer for the agentic web is a large, largely unsolved problem.
One example of a major challenge is identity and access management for agents deployed at the enterprise level. Agents can access company data and send documents to clients. You can see how this can go wrong easily. Inisight partners published a small landscape on this particularly:
The Israeli Angle
Israel’s startup ecosystem has a long track record of building foundational infrastructure layers: cybersecurity, cloud networking, developer tools. The agent web is a natural fit.
Tavily, founded in late 2023 by Rotem Weiss, Assaf Elovic and Yuval Rozio, built what may be the most quietly important piece of early agent infrastructure: a search engine designed specifically for AI agents to retrieve clean, structured, real-time data, not for humans to browse. The company served IBM, Cohere, and Groq and powered over one million developers before being acquired by Nebius for up to $400 million in February 2026, roughly 18 months after founding. Arguably the fastest significant exit in Israel’s recent AI history.
Onyx Security emerged from stealth this month with $40 million in funding, focused on securing and governing AI agents operating inside enterprises, a category that will only grow as autonomous systems move from experiments to production workflows.
Anchor Browser is building a cloud-hosted browser that lets AI agents interact with websites in secure, authenticated environments — solving one of the core practical problems of agent deployment at scale.
Together, these companies are addressing different layers of the same thesis: if agents become real, economic users of the web, they’ll need search, execution, security, and control planes built for them. The technical depth and threat-aware mindset that made Israel great at cybersecurity maps well to those problems.
Questions Every Founder Should Be Asking
If you’re building a product today, in any category, there’s a short checklist worth working through honestly:
- Is your product agent-accessible? Can an AI agent interact with your service programmatically, without rendering a full UI? Do you expose clean structured endpoints? Do you have an MCP server? If not, you may be building something that needs significant rearchitecting as agentic traffic grows.
- Are you building a moat in agent-specific infrastructure? The picks-and-shovels opportunity here is large and genuinely early. Agent identity, agent billing and rate-limiting, agent analytics, agent memory, agent reputation, agent-native storefronts, these are real infrastructure categories without clear winners yet. We invested in Aigency AI which is smack in this space.
- Have you thought about your agent pricing model? If agents can execute tasks that previously required dozens of human sessions, your per-seat or per-click pricing may not hold. Consumption-based or outcome-based pricing designed for agent velocity could become a significant competitive advantage.
- Are you building agent-native or just agent-compatible? There’s a difference. Agent-native means clean APIs, structured outputs, machine-readable content by default, no dark patterns. Products built from the ground up with agents as first-class users will compound their advantage as agentic traffic grows. The winners may not look like prettier websites. They may look like programmable surfaces, trusted endpoints, and machine-first APIs with human oversight layered on top.
- The opportunity is that entire categories are still wide open. Anthropic’s recent research on AI agents shows that 49.7% of agent tool calls are in software engineering. The next category, back office automation, is at 9.1%. Most other domains sit below 5%. Healthcare is at 1.0%, legal at 0.9%, finance and accounting at 4.0%. Many industry verticals barely register in current agent deployment.
We are moving from a web of pages to a web of actions. The first era of the internet was about publishing information for humans. The second was about platforms aggregating human attention. The next era looks more like an execution layer for agents, and when that shift happens, distribution changes, monetisation changes, defensibility changes, and the very meaning of a “user” changes.
Adobe Analytics reported a 4,700% year-over-year increase in traffic from AI agents to US retail sites in a single month in 2025. Visa, Mastercard, PayPal, and Google have all launched protocols to enable agents to transact on behalf of users. The companies that figure out how to serve, monetise, and govern agent interactions, rather than only human ones, will own a significant share of the next decade of the web economy.
If you’re an Israeli founder building in this direction, whether around agent infrastructure, trust, payments, security, or agent-native consumer experiences, I’d love to hear from you. At Remagine Ventures, we’re actively looking to speak with founders at the very beginning of their journey.
The bots are coming. The question is who builds the world they’ll inhabit.
I'm a former general partner at google ventures, head of Google for Entrepreneurs in Europe and founding head of Campus London, Google's first physical hub for startups.
I'm also the founder of Techbikers, a non-profit bringing together the startup ecosystem on cycling challenges in support of Room to Read. Since inception in 2012 we've built 11 schools and 50 libraries in the developing world.
- The Web’s Next Customer Isn’t Human - March 13, 2026
- Weekly Firgun Newsletter – March 13 2026 - March 13, 2026
- In 2026 Gaming companies must deal with AI supercycle and VC winter to find new moats - March 12, 2026
