Agent Readiness Scanner Update: Noise Floor, Harness Lift and a Commerce Task

The report in the agent readiness scanner has been rebuilt to cut repetition and surface what to fix first. The Live Agent Run now publishes its own noise floor so small framework differences read as noise. It also prints harness lift, which is what it adds over a plain model. Sites that already show a payment surface get an eighth question about where an agent could pay.

If you are deciding whether your site is usable by AI agents, the basic scan is unchanged and still free with no signup.

Why one agent run was never enough

Agents do not give the same answer twice. With one run per framework you cannot tell whether a score gap is real or random, and a site that returns 73 to 92 across five frameworks on a 0 to 100 scale could be showing a stronger harness or simply luck on that single run.

The old report made that harder to judge. It repeated the same findings three times in three different formats, so every issue appeared in three places and readers scrolled through three screens of restatement before they could fix anything. The duplication buried priority under repetition.

The control is now the error bar

The control is a plain model with one fetch tool and no agent harness. That is what most teams ship first when they give an agent access to the web.

It now runs three times on identical input. The report publishes the spread between its best and worst run and states the rule beside the number. Any difference between frameworks that falls inside that spread is noise, not signal.

The report also prints harness lift. It is the average framework score minus the control median, and it shows what the harness adds on this specific site. Without the control you cannot tell whether a low score is a framework limit or a site that agents cannot read.

A payment question, only where a payment surface exists

Some sites already expose a payment surface. The deterministic scan proves it exists when it finds an x402 endpoint, AP2 or ACP discovery, or a payment link.

When that surface exists the Live Agent Run adds one question to the usual seven. It asks for the URL where an agent could pay without a human, so the test moves from existence to findability. Sites without a payment surface are never asked that question and are never marked down for it.

A report that says everything once

The report now opens with the overall score and all six category gauges in one row, followed by What to fix first. Failing checks are ordered by weight and labelled High impact, Medium or Low, and each carries a ready-to-paste fix prompt for an AI coding agent such as Claude Code or Cursor.

Counts reconcile in one line. A recent scan of linear.app read: 18 need work, 16 pass, 6 do not apply to this site.

The report also states the cost of format. One scanned site served 39,409 tokens as HTML and 3,064 tokens as markdown, a 92% saving for every agent that reads it when markdown is available.

Run the agent readiness scanner on your own site. The HTTP scan is free with no signup. An account includes three Live Agent Runs per month. Daily monitoring is available for sites that need to stay agent-ready.