The Agent Internet will not be another place businesses have to enter. It’s a new layer across the internet we already have, one in which machines increasingly participate in discovery, evaluation and transactions on behalf of people. Long promised, now here. So how do we begin to make sense of it?

A few years on from the word ‘metaverse’ becoming commonplace, you would be forgiven for lacking enthusiasm for the next supposedly transformative version of the internet. The metaverse remains niche, often relies on specialist hardware, and many of the much-hyped applications never gained meaningful traction. Apparently persuading people to inhabit an entirely different internet was a bigger problem than many expected.

The emerging Agent Internet takes a very different route.

New layers of the web

The metaverse required consumers to adopt a new destination, new behaviours and often new hardware. The agentic shift doesn’t. AI agents can operate across websites, APIs, search engines, ecommerce platforms, SaaS tools and existing business systems.

The change is therefore less about where we go online and more about who — or what — does the navigating. Increasingly, we can give an agent an objective and let it research, compare, communicate and act on our behalf. That has implications far beyond AI chat.

From chats to agents

The change has happened remarkably quickly. ChatGPT familiarised millions of people with conversational AI in 2022. We have already moved from asking an AI for an answer towards giving it a task.

Projects such as OpenClaw showed how assistants could persist across messaging services, use tools and run multiple isolated agents. Hermes Agent combines persistent memory with the ability to delegate work to subagents.

Now those ideas are becoming mainstream products. Grok Bot offers always-on agents with their own computers that can work across applications. OpenAI’s Dots similarly pushes towards persistent agents that continue working beyond an individual chat, while Meta’s Muse can handle tasks across connected services and is being extended into hardware through Muse Charm.

The real shift is from one-off conversations to persistent digital workers, and increasingly to agents coordinating with other agents.

A second audience for the internet

This becomes particularly interesting when agents act as customers, researchers or buyers.

The traditional digital journey starts with a search, followed by links, website visits, comparison and eventually a purchase. AI can collapse much of that into a short conversation. Adobe found that referrals from generative AI tools to US retail sites increased 693.4% year-on-year during the 2025 holiday season. The base remains modest, but the direction is striking.

A human may respond to branding, photography, storytelling or reputation. An agent may care more about price, specifications, availability, reviews, compatibility, contractual terms and credible evidence. Businesses may therefore need two digital shop windows: one designed to persuade people and another structured so machines can reliably understand what is being offered.

Accenture found that 56% of consumers would instruct an agent which brands to consider, while 37% of “behaviorally loyal” consumers would allow one to switch from a preferred brand if it found a better fit.

This is not just a marketing issue. Procurement, recruitment, travel, finance, customer support and software purchasing could all involve agents screening options before a person becomes directly involved.

From tools to an agent network

The infrastructure for this is also beginning to form. Model Context Protocol (MCP) gives agents a standard way to connect to tools and data, while Agent2Agent (A2A) is designed so agents from different systems can discover one another, delegate tasks and collaborate. Anthropic reported more than 10,000 active public MCP servers by December 2025.

That moves us towards something more significant than better assistants: an internet in which software actors increasingly transact and coordinate across organisational boundaries.

For businesses, that creates opportunities to automate research, operations, customer service, procurement and routine administration. Meta’s newly announced Muse for Small Business, for example, connects with tools including QuickBooks, Shopify, Slack, Stripe, Notion and Canva, while keeping publishing, sending and spending subject to user approval.

But autonomy brings friction

This future should not be treated as inevitable or frictionless. Accenture found that 74% of consumers would delegate routine tasks to an agent, but only 9% were open to fully autonomous purchasing.

Even businesses already experimenting with agents are finding limits. Gartner reported in 2025 that half of surveyed martech leaders lacked the technical and data-stack readiness required for deployment, while 45% of those piloting or implementing vendor agents said performance fell short of expectations.

Security, permissions, incorrect actions, accountability and trust all become more important when software can act rather than simply advise.

What should businesses do now?

The sensible response is experimentation, not wholesale reinvention.

Identify repeatable workflows where agents can genuinely save time. Make important business information structured and accessible. Decide what an agent should be allowed to read, change, send or spend. Build approval points around consequential actions. And start considering how your organisation looks not only to human customers and partners, but to their software representatives.

The metaverse asked us to visit a different internet. The Agent Internet may prove far more consequential because it changes who uses the internet we already have.