From Website to Controlled Agent Interface
A website can remain the human experience while a parallel machine channel serves structured, current, policy-controlled business data.
Most businesses do not need to replace their website to support AI agents. They need a second interface with different operating properties.
The human site is designed for persuasion, navigation, accessibility, and brand expression. An agent-facing interface is designed for stable identifiers, explicit schemas, current facts, authorization, and predictable errors. Asking one surface to do both jobs usually produces a brittle compromise.
Start with source authority
The first architectural decision is not which protocol to adopt. It is which system is authoritative for each fact. Product availability may belong to an inventory API. Contract terms may belong to a controlled document. Company positioning may belong to a maintained page. A gateway should preserve those distinctions instead of flattening every source into an undifferentiated index.
Canonical records need provenance and an observed time. When sources disagree, the conflict should remain visible until an explicit authority rule selects the published value. Source order is not a defensible conflict-resolution policy.
Publish narrow interfaces
Markdown is useful for readable context. REST is useful for conventional integrations. MCP provides a standard tool and resource surface for compatible agent hosts. Commerce profiles such as UCP can advertise supported capabilities. These interfaces should be projections of the same canonical data, not separately maintained copies.
Every published capability also needs a boundary. Public brand context may require no credential. Real-time inventory may require a scoped client. A quote request may require an approval step. The interface format does not decide that policy; the business does.
Operate the channel
An agent interface is production infrastructure. Teams need to monitor source freshness, failed ingestion, response latency, authorization failures, tool versions, approvals, and billed units. They also need a safe way to turn a capability off without taking down the human website.
The practical goal is modest: give machines a more reliable path than screen scraping, while keeping business owners in control of source authority and execution.