AgentDrives Agent
AgentDrives Agent is the AI agent of AgentDrives (agentdrives.com). It test-drives
software products the way an AI agent would: it reads the docs, tries to sign up, get an API key, use the product, check
how it would pay, cancel and delete its account. AgentDrives publishes how far it got, with the evidence;
owners can reply next to the result and order a new drive.
It runs on our server in a data center in Germany, not in anyone's browser at home, and says who it is: its HTTP
client and its browser name it and sign every request. It works over a product's API, CLI or MCP first, and opens its
browser, on the same server, only for steps that need one, such as a sign-up form or a checkout it walks up to the pay
button.
How it behaves
- One careful user.
One drive per product at a time, at a person's pace (about one page action every few seconds; reading a product's public docs before a drive, one page at a time, about one a second); up to three drives at once across different products, about a hundred a day. A complete drive (every Drive and Audit an owner orders, and every full drive we start ourselves) may send one short burst of 10–20 calls to one endpoint to check rate limiting, and stops at the first 429. Its own test accounts, its own test inboxes, its own GitHub and Google accounts for their sign-in buttons (it allows only what sign-in needs and takes the access back at the end), no load tests, no messages to real people, no spending beyond the one payment below.
- Payment.
It never buys a plan on a drive we start ourselves. Where a product takes payments from agents on its own endpoint (x402 or MPP),
the drive may pay it once: at most $1, from our own wallet. It buys a paid plan only with a one-use code or a test-mode
login its owner gives, or when the owner adds the plan to an Audit, charged to them at cost.
- Stress test, on every complete drive.
Every Drive and Audit an owner orders, and every full drive we start ourselves,
also tests the unhappy paths on our own test account: a few malformed requests, a wrong key of its own, one create sent twice,
a burst of 10 to 20 reads to one endpoint, and its core tasks run three times. It stops at the first 429. Shorter drives
(up to 45 minutes: re-checks, and products driven as part of a comparison) skip it.
- Not a crawler.
It visits what a user would to sign up and use the product, and before a drive it reads the product's
public docs (the pages its llms.txt, sitemap and menus list, at most 600, obeying robots.txt) to judge them. It does not index sites, does not
train models on what it reads, and keeps only what a report needs as evidence (screenshots and short quotes of the steps it took).
- Monthly survey.
Once a month it reads a few public pages of popular software sites, such as the home page,
robots.txt, llms.txt and the sign-up page, and we publish totals only. It obeys robots.txt, signs every request and asks
each site at most once every 3 seconds. Blocking AgentDrivesAgent stops it too.
- robots.txt.
Before a drive it reads the product's robots.txt. If the product's address is disallowed for
AgentDrivesAgent (or for all user agents), it does not drive that product, unless the owner ordered the drive.
So robots.txt is the way to opt out of editorial drives (drives we start ourselves). During a drive it follows robots.txt for every page it reads. Inside the test account it creates (the sign-up form, the verification link, logging in, the key page, calls with its own key, cancel and delete), a rule for all user agents is a rule for crawlers, so it goes on. A rule that names AgentDrivesAgent, or AI agents that act for a user, always stops it.
- Bot checks.
It never solves, skips or works around a CAPTCHA or bot check, never uses a solving service,
never retries around a block (403, 429 or a challenge) and never hides that it is automated. A bot check stops the drive at
that step, and the report says so. In some drives a person from our team then does that one step by hand, in the agent's
browser or their own, and the agent goes on from there; the report marks every step a person did as done by a human,
never by the agent, and the agent's level counts only what it did on its own.
- Opt-out is honoured.
Block AgentDrivesAgent in robots.txt, or write to us, and we stop driving
your product. Results already published stay, with their date and your reply if you send one.
How to recognise it
- User-Agent (HTTP)
AgentDrivesAgent/1.0 (+https://agent.agentdrives.com/bot)- User-Agent (browser)
- Chrome's string followed by
AgentDrivesAgent/1.0 (+https://agent.agentdrives.com/bot)
- robots.txt token
AgentDrivesAgent- Signature-Agent
"https://agent.agentdrives.com"- Egress addresses
167.233.145.55, 2a01:4f8:1c16:f310::1 (for information; not exclusive to the agent)
Its browser and its HTTP client sign every request with
Web Bot Auth (HTTP Message Signatures, RFC 9421, Ed25519,
tag="web-bot-auth"). The signature covers @authority and signature-agent and is valid
for 60 seconds. Not signed: calls a product's own SDK, CLI or MCP client makes (they go out as that client sends them),
and reading public docs through a search or fetch tool. Trust the signature, not the User-Agent or the address: anyone
can copy those.
Verify, allow or block it
- Bot managers.
Cloudflare, Vercel, Akamai, HUMAN, DataDome and other bot managers that support Web Bot Auth verify the
signature against our key directory. To let it through, allow verified agents (or this agent by name) in your bot settings.
- Your own code.
Fetch the key directory, check that the Signature-Agent header is
"https://agent.agentdrives.com", and verify the signature with any RFC 9421 library, e.g.
cloudflare/web-bot-auth.
- Block it.
Add to robots.txt, or write to us:
User-agent: AgentDrivesAgent
Disallow: /
Operator and contact
AgentDrives, the Netherlands. Questions, corrections, abuse reports or an opt-out:
hello@agentdrives.com.