The Agent Traffic Thesis
A prediction: within two years, more product purchases, API calls, data lookups, and form submissions on the web will be initiated by AI agents than by humans directly.
This is not a prediction about AGI. It's a prediction about economics.
The Cost Differential
A human browsing a website to compare flight prices spends 15-30 minutes, processes information visually, and makes one purchase. An agent performing the same task takes seconds, processes structured data, and can run in parallel across dozens of providers.
The cost per transaction for agent-mediated tasks is already 10-100x lower than human-mediated tasks when the infrastructure supports it. The infrastructure is what's catching up.
Where It Starts
Agent traffic won't replace human browsing uniformly. It will dominate transactional use cases first:
- Price comparison and purchasing
- Form submissions (applications, registrations)
- Data aggregation and monitoring
- Appointment booking
- Recurring administrative tasks
These are high-frequency, low-creativity tasks where speed and cost matter more than human judgment. The human becomes the decision-maker; the agent becomes the executor.
What Blocks It
Two bottlenecks:
- Authentication — most web services still assume a human with a browser session. Agent identity and auth standards (like x402 for payments) are nascent.
- Structured interfaces — agents interacting with websites designed for humans waste tokens on DOM parsing and screenshot analysis. WebMCP directly addresses this.
Remove these bottlenecks and the economic pressure is entirely in one direction: more agents, fewer direct human interactions with transactional web services.
The Counter-Argument
Some argue that humans will always prefer direct control over important transactions. This is true for high-stakes, low-frequency decisions (buying a house). It is not true for routine transactions where the human's preferences can be specified in advance.
The question isn't whether humans can do these tasks. It's whether they will, when an agent can do it faster, cheaper, and without errors.