The Agent Internet
The internet is adding a new kind of participant. For thirty years, the web has been a two-party system: humans and the services they visit. You open a browser, go to a website, and do things. That's changing. Artificial Intelligence (AI) agents can now do many of those things on your behalf, handling the browsing, comparing, clicking, and form-filling. You get the result without doing the work.
This sounds incremental. It isn't. When the participant changes from a human to software acting for a human, almost everything about how the internet works needs to be rethought: how services get discovered, how they get paid, how trust is established, how safety is maintained, and how power is distributed.
This series of papers works through those problems. Together, they outline the emerging Agent Internet. Arcede has built the first Agent Internet Runtime (A.I.R.) which will play a critical role in this transition.
The Core Idea
Intelligent compute is becoming a utility. Electricity went from a novelty to infrastructure over about fifty years. Computing followed a similar arc. Now computing intelligence, the ability of software to reason, plan, and act, is following the same path. The cost of running an AI model has dropped roughly tenfold every year since 2023. Computing intelligence is becoming something you meter and consume, like electricity or bandwidth.
When computing intelligence is a utility, it creates demand for a new kind of infrastructure. You don't just need servers and bandwidth anymore. You need systems that let intelligent software interact with the internet on behalf of humans, safely, efficiently, and economically. That infrastructure is what we call the agent internet. These papers describe what it requires.
What Changes
Discovery. Humans find services by searching, browsing, and following links. Agents don't browse. They match what the user wants against what services offer. This means services need to declare their capabilities in a machine-readable format. The competitive advantage shifts from marketing and search optimization to actual service quality.
Economics. The two-party web is funded by advertising: services capture human attention, sell that attention to advertisers, and the user pays with their time. Agents don't have attention. They skip the ads, ignore the engagement tricks, and go straight to the task. This breaks the advertising model and replaces it with a completion economy, where services get paid when they successfully complete a task, not when they capture eyeballs.
Efficiency. Every agent action costs money: AI inference, network requests, computation. When the user is an agent, the cost is metered compute. This makes efficiency a direct competitive advantage: the service that delivers results faster and with fewer computational steps is cheaper and gets more traffic.
Trust. A three-party system (human, agent, service) has fundamentally different trust requirements. The human needs to trust the agent. The service needs to verify the agent is authorized. The agent needs to prove who it represents. None of the existing web's trust mechanisms were designed for this.
Safety. An agent that can act on your behalf can also act against your interests, by making mistakes, exceeding its authority, or being manipulated. Supervision, graduated human oversight of agent actions, is what makes delegation viable. Without it, no one would trust an agent with anything that matters.
Risk. Agents introduce failure modes that don't exist on the two-party web: systematic errors affecting thousands of users simultaneously, market concentration from algorithmic routing, the hollowing-out of information sources, and the gradual drift toward unsupervised autonomy. These are foreseeable and addressable, but only if the builders confront them directly.
The Infrastructure Gap
Even with all of the above solved, agents still hit three walls:
- They can't pay. An agent that finds the best deal can't complete the purchase without payment infrastructure designed for software, not humans.
- They can't prove who they are. An agent acting on your behalf can't authenticate using your passwords. It needs a new kind of identity that proves authorization without holding credentials.
- They can't communicate. An agent that can research and pay still can't send an email, schedule a meeting, or negotiate with another agent on your behalf.
These three — payment, identity, communication — are meta-capabilities. They don't unlock one service; they unlock every service. Each must be built as an open, competitive ecosystem, because a monopoly over any of them would give a single entity control over what agents can do, see, and say.
The Foundational Series
Economics & Operations
The Agency Problem
Infrastructure
Who This Is For
- Builders making agent products or infrastructure.
- Service providers wondering how to participate in the agent economy.
- Business leaders evaluating the commercial implications of agent adoption.
- Policymakers thinking about AI regulation.
- Anyone curious about how the internet changes when software starts using it on our behalf.