Your users' AI agents are failing on your product. We know where.

Why

Agents use your software now. They fail quietly.

A dead reference page, an inert parameter, a 500 that blames the target site. A human grumbles and works around it. An agent burns your credits, gives up, and picks the next vendor. No ticket, no signal, just a trial that went quiet.

Method

Four stages. The only input is what a stranger can find.

1. 01Map. Your claimed capabilities, reconciled against the whole field.
2. 02Build blind. A fresh agent, your public docs, a key we bought. No hints.
3. 03Score. Four difficulty tiers, graded 2 / 1 / 0 against an oracle fixed before the run.
4. 04Evidence. Every finding ties to a stored request and its response. No payload, no finding.

Findings

Four classes. One of them is good news.

A · Behavioural

The API did something the docs never said it would.

B · Absence

Something blocked the agent and was never explained.

C · Documentation

Unreadable docs, or docs that describe what isn't there.

D · Capability

Strength you are not getting credit for.

What you get

Your failures, and where they put you in the field.