LAST MILE Where agents actually land

The playbook

Every best practice, end to end.

The Index shows where the route breaks. Each best practice below reads as one causal chain — the failure mode an agent hits, what you change, and what the agent can then do — grouped by the dimension it repairs and ordered by tier.

Best practices
7
Performance areas covered
6
Remediation tiers
3
Evidence
versioned, cited
Input affordances Controls accept agent input. Failure transparency Errors explain the recovery. Consent wall Blocking gate measured.
Tier 0Remove the blocker
Tier 1Publish the facts
Tier 2Offer the tool
Verified route restored Agent journey completed.

Find a best practice.

Filter the playbook by the performance area, maturity level, or operating status you need.

Showing 7 best practices.

Discoverability

2 best practices

Do agents find you?

Measured by: Discoverability

Open the Index →
Tier 1 · Publish the facts Production

Add semantic markup and structured navigation to expose entry points

Mark up navigation landmarks with ARIA roles, use descriptive link text matching common task keywords, and add structured data (JSON-LD) so agents identify the right entry point without guessing.

Failure mode Ambiguous entry points

Generic links force agents to crawl and guess

click here · learn more
What you change Structured navigation

Landmarks and descriptive links expose the task

<nav aria-label="Primary">
What the agent can do
Agents reach the right capability with less search.
Performance area
Discoverability
Tier 1 · Publish the facts Emerging

Publish an llms.txt capability index

Serve /llms.txt describing the site's key pages and capabilities in Markdown so agents can orient without crawling.

Failure mode Crawl to orient

Capabilities are scattered across the site

crawl → infer
What you change Capability index

Key pages and functions are named in one place

/llms.txt
What the agent can do
Agents can orient before entering the interface.
Performance area
Discoverability

Understanding

3 best practices

Can agents read & navigate you?

Measured by: Structured path · Input affordances · Failure transparency

Open the Index →
Tier 0 · Remove the blocker Production

Publish core info as agent-readable content

Ensure the answer to the task is available as readable static content (a text table, a paragraph) so an agent can extract it without navigating a transactional flow.

Failure mode Answer behind interaction

The fact exists only after a brittle flow

answer → hidden state
What you change Readable fact

The answer is present as stable page content

text · table · paragraph
What the agent can do
The deterministic check can verify the answer.
Performance area
Task completion
Tier 0 · Remove the blocker Production

Replace hostile date widgets with agent-readable inputs

Remove readonly/hidden date fields and custom datepicker overlays that reject programmatic input; use a standard <input type="date"> or publish the fact as readable text so an agent never needs to fight the widget.

Failure mode Hostile date widget

Overlay rejects programmatic input

readonly + hidden input
What you change Agent-readable input

A native control accepts a deterministic value

<input type="date">
What the agent can do
Agents can enter dates without fighting the widget.
Performance area
Input affordances
Tier 0 · Remove the blocker Production

Surface real validation errors instead of silent dead-ends

When a form submission fails, return a visible, machine-readable error message — role="alert" or aria-invalid — so the agent knows the attempt failed and why.

Failure mode Silent dead end

Submit fails without a usable explanation

submit → no response
What you change Visible validation error

The field and correction are surfaced together

role="alert"
What the agent can do
Agents know what failed and how to recover.
Performance area
Failure transparency

Interaction

2 best practices

Can agents operate you? (unpaid)

Measured by: Consent wall · Structured path · Task completion

Open the Index →
Tier 2 · Offer the tool Emerging

Expose WebMCP tools for true agent workflows

Implement window.__webmcp (or a link[rel='mcp'] manifest) exposing typed tool definitions for core tasks (search, booking) so agents invoke structured operations instead of driving the DOM blindly.

Failure mode Blind DOM driving

The agent infers controls from presentation

inspect → guess → retry
What you change Typed agent tool

A declared operation returns structured output

tool(searchFlights)
What the agent can do
Agents invoke a supported path instead of guessing.
Performance area
Structured path

Transaction is tracked as an outlook on the methodology page →

Find the break See which airline performance areas fail in verified agent runs. Open the Index → Verify the method See how performance areas, access gates, and outcomes are scored. Read the methodology →