LAST MILE Where agents actually land

Methodology

A real agent drives. The harness decides.

Last Mile measures whether browser agents can find, understand, and operate a site through a real pre-purchase journey. The agent chooses the actions; it never grades itself.

Three gates. One journey.

A site can surface in search, expose readable information, and still fail when an agent must operate it. Each gate is tested separately so the Index shows where the journey actually ends.

  1. Discoverability

    Do agents find you?

    Can an agent surface the airline and reach the capability the task requires?

    Inspect gate evidence

    What decides this gateA real agent is prompted for the task — does the airline (or its function) surface and get reached at all? If not, the journey never starts.

    8 standards and protocols tracked

    • robots.txt established

      Tells crawlers and AI bots which paths they may fetch. · IETF · RFC 9309

    • Sitemap established

      An XML list of your URLs so crawlers can find every page. · sitemaps.org

    • Link headers established

      Declares related resources in the HTTP Link header — no HTML parse needed. · IETF · RFC 8288

    • DNS-AID experimental

      Publish SVCB DNS records so agents discover your services before any HTTP call. Early / experimental. Impossible on unsigned zones like *.workers.dev. · IETF draft · community N/A on *.workers.dev (unsigned zone) — record as N/A, never as a failure.

    • MCP Server Card experimental

      A discovery card advertising the MCP tools your site exposes. Very early — few sites use it. · Anthropic · MCP (SEP-1649)

    • API Catalog emerging

      A machine-readable index of your APIs at /.well-known/api-catalog. · IETF · RFC 9727

    • OAuth discovery established

      Lets agents auto-discover your authorization server's endpoints. · IETF · RFC 8414

    • Web Bot Auth experimental

      Agents cryptographically sign requests (built on RFC 9421) so you can verify and allow trusted bots. Draft, but shipping. · Cloudflare + Google · IETF draft

  2. Understanding

    Can agents read & navigate you?

    Can it interpret offers, rules, inputs, and errors without fighting opaque UI?

    Inspect gate evidence

    What decides this gateCan the agent parse offers and read the inputs and errors, or is content locked in JS / PDFs? Structured path, input legibility and error legibility are the tells.

    3 standards and protocols tracked

    • Markdown negotiation emerging

      Serve clean Markdown when an agent sends Accept: text/markdown (~80% fewer tokens). Beta, ~4% adoption. · Cloudflare · HTTP content negotiation

    • schema.org / JSON-LD established

      Structured-data vocabulary that makes offers, fares and policies machine-readable. Established. · schema.org · W3C (JSON-LD)

    • NDC established

      Airline XML standard for offers and orders — most carriers already have it. · IATA

  3. Interaction

    Can agents operate you? (unpaid)

    Can it search, select, and recover from errors without a human taking over?

    Inspect gate evidence

    What decides this gateDoes the agent complete a search, check-in or selection end to end — recovering from errors — with no human handoff? Unpaid task completion and failure recovery.

    3 standards and protocols tracked

    • WebMCP experimental

      Web pages expose typed tools a browser agent calls directly, skipping the DOM. Experimental — Chrome origin trial, Gemini-only. · Google & Microsoft · W3C CG

    • MCP (actions) emerging

      Open protocol connecting agents to your tools and actions. Broadly adopted; ~8 months old. · Anthropic

    • OAuth Protected Resource established

      Metadata telling agents how to authenticate to your protected APIs. · IETF · RFC 9728

One public grading language Fails, Struggles, and Completes summarize measurable run records. The underlying harness contract remains pass, fail, or blocked.

  • Fails

    Measured tasks do not reach a verified completion.

  • Struggles

    Some tasks complete, but material friction remains.

  • Completes

    The defined task set reaches verified outcomes.

The benchmark follows the purchase journey.

The first airline sweep prioritizes the tasks with the clearest commercial impact: helping a traveler choose a flight and reach the final commitment point without staged credentials.

  1. Inspiration & planning

    Destinations, flexible dates, entry requirements, and policy questions.

  2. Compare & select

    Flight search, fare comparison, total cost, baggage, and connection risk.

  3. Book to the wall

    Bundles, ancillaries, passenger details, and the full order review before payment.

The measured run stops at the purchase or commit wall. Payment itself is not attempted.

One proof trail from task to published result.

Static readiness signals stay visible, but they never substitute for the browser-agent run. Every published claim resolves back to observable task evidence.

  1. 01

    Issue the task

    A real customer request is paired with a known, deterministic success check.

  2. 02

    Drive the site

    A browser agent chooses actions while Tarmac records the trace and friction signals.

  3. 03

    Verify the result

    The harness checks observable evidence and returns pass, fail, or blocked.

  4. 04

    Publish the evidence

    Persisted run records generate the Index, site dossier, and report artifact.

Future scope

Transaction begins beyond the money line.

The agent completes money-moving actions on the customer's behalf — with payment, agent auth and trust. Depends on Interaction. Below the money line. It remains the highest bar, but it is not part of the current airline Index. Today, Last Mile reports readiness up to payment, not payment itself.

Payment protocols we track 6 emerging standards kept in view View protocols Hide protocols
  • ACP emerging

    Agentic Commerce Protocol — open checkout standard any ACP agent can pay. Beta; powers ChatGPT Instant Checkout.

  • UCP emerging

    Universal Commerce Protocol — merchants declare commerce capabilities agents discover and transact. Live in Google AI Mode / Gemini.

  • x402 experimental

    Revives HTTP 402 for instant programmatic (stablecoin) payments. Young and volatile — optionality, not a baseline.

  • MPP experimental

    Machine Payments Protocol — multi-rail agent payments over HTTP 402 (Stripe Agentic Commerce Suite). New.

  • AP2 emerging

    Agent Payments Protocol — agents authorize payments via a cryptographically signed Mandate. Composes with UCP.

  • NDC OrderCreate established

    The NDC message that creates a booking / order programmatically.

Blocked is an access boundary, not a failed journey.

Tasks are graded by how far an agent can reach without staged credentials. A gated task that reaches its expected wall is recorded as blocked and excluded from measurable Index math.

Pass
The deterministic success check found the required evidence.
Fail
The measurable journey ended without a verified answer.
Blocked
An honest credential or access wall stopped the run; it is excluded.
Access tier Agent reach How far the browser agent can proceed before it meets a legitimate access boundary. Measured today
Public End to end The task is public, so the agent can attempt it and verify the result without staged credentials. Included
Booking flow To payment wall The agent can navigate the booking flow through the final review, then stops before money moves. Included
Booking ref Stops at booking lookup A real booking reference and passenger name are required before the agent can continue. Not yet scored
Login Stops at sign-in A staged customer or frequent-flyer account is required before the task can be measured. Not yet scored
Eligibility Stops at eligibility check The flow needs a real disrupted booking or another eligibility condition that cannot be staged reliably. Not yet scored
Hybrid Varies by flow Part of the journey is public, but another step requires credentials or case-specific data. Not yet scored
Staff Not customer-facing This task belongs to staff systems, so it is outside the public website benchmark. Not yet scored
Controlled proof See the same task set move from failing to completing. Open the Playground → Published evidence Read the reports generated from persisted run records. Browse reports →

Does the driving model matter? Inspect the cross-model sweep →