AF-0 to AF-5 comparison

From a hard-to-interpret website to an agent-native platform.

Explore restaurant, municipality and Tokenizart examples. Each stage shows how an answer changes as better sources, data and tools appear.

The comparison is illustrative: better evidence improves answer precision, but does not guarantee indexing, ranking, recommendation or identical wording in GPT, Gemini, Claude or another model.

Progress is not automatic.

Agentic transformation happens in layers.

Each level requires public evidence, real implementation, validation, explicit limits, human approval where needed and a new measurement.

AF-0

Invisible

The agent cannot find enough evidence.

AF-1

Discoverable

Coherent robots, sitemap and public routes.

AF-2

Understandable

Answers, structure, sources and dates.

AF-3

Tool-enabled

Verifiable public contracts and resources.

AF-4

Delegable

Identity, permissions, consent and audit.

AF-5

Agent-native

Coordination and payments only where they add value.

Case
Observed maturity
Restaurant

Is it open today, does it have a gluten-free menu and can I book for eight people?

Before, or without enough evidence

I cannot find sufficiently clear and current information to confirm hours, menu or reservations.

Illustrative answerAF-2 ยท Understandable

It is listed as open from 12:00 to 23:00 and declares gluten-free options; availability still requires confirmation.

Answers, structure, sources and dates.
  • Updated hours, address and official channels.
  • Structured menu with allergens, prices and validity.
Next move

Audit evidence before promising capabilities.

The score changes only when an improvement can be observed at the public origin. MCP, A2A, WebMCP or x402 do not count merely because they appear on a roadmap.

Audit a website