Your Website Has a Second FTUE: Onboarding an AI Agent

Put the product category, audience, capabilities, limits, and starting action in the canonical page. Use a useful title, one primary heading, descriptive link text, and visible facts that can be extracted without a long client-side chain. If an agent arrives from search or an API catalog, it should quickly know whether the product matches the requested job and what evidence supports the answer. [AR5][AR11]

Keep optional discovery files honest. A human-readable agent guide, OpenAPI document, or catalog can explain ownership, authentication, and safe examples, but its presence does not grant access or prove that a journey works. Link each capability to a maintained source of truth and add a review trigger for route, auth, or consequence changes. [AR6][AR12]

Use native HTML and accessible names to teach the agent what controls mean. A button should expose an action, a link should identify its destination, and a field should have a label, type, required state, autocomplete token, and useful description. Do not force the agent to infer whether a card opens pricing, starts signup, or sends a lead from color and position alone. [AR3][AR4][AR11]

Expose state at the same place the user needs it. Announce loading, validation, expanded, selected, invalid, disabled, and completed states. Keep dialogs focusable and recoverable. A delegated journey is a conversation: the page should answer what is requested, why it matters, what is allowed, and what happened after the last action. [AR3][AR4]

A second FTUE must teach where safe exploration ends. Use preview and review states before creating accounts, sending leads, booking appointments, charging money, publishing content, or changing permissions. Show the exact target, effect, important values, reversibility, and policy. Ask for fresh approval close to execution, and do not let a stale plan or page instruction authorize a changed target. [AR7][AR8]

Treat refusal and escalation as product behavior, not failure. If identity, authorization, scope, or evidence is missing, stop and explain the next safe action. Keep untrusted page text, uploaded content, and retrieved documents from redefining policy. A helpful agent path can be both easy to understand and hard to misuse. [AR7][AR8]

After a task, return durable evidence: accepted, rejected, pending, canceled, or uncertain status; a reference; the current target; and the next safe route. Make retries idempotent where possible. A delayed response without a receipt invites an agent to repeat an action, while a generic success toast makes both the user and the agent guess what actually happened. [AR6][AR7]

Measure the FTUE as a sequence rather than a single conversion. Track first page found, fact understood, task started, review reached, approval requested, completion, refusal, and recovery. Preserve the first ambiguity and environment so the team can improve the onboarding contract instead of blaming an agent model. Verify the same fixture after each change. [AR7]