AI & Agents

Search retrieves. Agents act.

Retrieval can tolerate ambiguity — a wrong result costs a click. Action cannot: an agent operating on stale or unresolved information produces a real-world failure.
01Two kinds of request

Traditional search

“HVAC companies in Duluth”

Ten blue links. A person reads, compares, calls, and does the reconciliation work themselves.

Agent request

“Find three verified HVAC companies serving my address that have availability tomorrow and can handle my specific system.”

No page answers this. It requires resolved entities, service areas, verification status, availability and equipment compatibility.

What that request depends on

  • Identity
  • Services
  • Geography
  • Availability
  • Pricing
  • Credentials
  • Trust
  • Permissions
  • Actions
  • Transactions
02The access path
  1. AI Agent
  2. Agent Interface / MCP
  3. Record Infrastructure
  4. Business · Place · Service · Offer · Event Records

Agents need a described interface, not a page to scrape.

An interface over a record can state what exists, what may be read, what may be changed and what actions are permitted — with a rate policy and an audit trail. Emerging protocols such as MCP are one attempt at this shape.

This is why we suspect structured record infrastructure becomes more important as agents take on more work, not less.

03Serving both internets

The Human Internet

  • Websites
  • Pages
  • Navigation
  • Brands
  • Stories
  • Images
  • Interfaces
  • Experiences

The Machine Internet

  • Records
  • Entities
  • Attributes
  • Relationships
  • Graphs
  • APIs
  • MCP
  • Citations
  • Permissions
  • Agents

The next internet must serve both.

Humans experience surfaces. Machines consume structured information. One representation can feed both.

Read the research →