---
title: "We Sent an Agent to Sign Up for 16 Products. Here Is Where It Stopped."
description: "Thirteen agent-first signup flows, run in September and again this week, plus five products through AgentMail's new AgentID. Real runs, product names left in, and where the agent needed a human."
date: 2026-10-13
author: Kat Laszlo
canonical: https://tansohq.com/blog/agent-signup-runs
---

> Full site context for agents: https://tansohq.com/llms.txt

# We Sent an Agent to Sign Up for 16 Products. Here Is Where It Stopped.

By [Kat Laszlo](https://www.linkedin.com/in/katrinalaszlo/) · October 13, 2026

Thirteen agent-first signup flows, run in September and again this week, plus five products through AgentMail's new AgentID, and where the agent needed a human.

Last week AgentMail launched [AgentID](https://www.agentid.com), a "Sign in with Google" for AI agents, and their newsletter opened with a problem I recognized. Their onboarding emails had started getting replies like "No idea what this is. My agent signed me up," and Browser Use had blocked their whole shared domain, because an email address was all it could see. I had been watching the same problem from the other side, because Tanso's Agent Usability test sends a real agent at a product and records where the agent gets stuck.

In September we ran that test against thirteen signup flows built for agents. After AgentID shipped, we went back on October 13 and ran the same thirteen flows again to see whether the September results still held, and we ran five products through the new sign-in button. This post covers what happened, with the product names left in, because a list that only shows the successes would not help anyone decide what to build.

## How the runs work

Each signup run gave a sandboxed Claude agent one task and nothing else: sign up as an agent, get a credential, and make one authenticated call. The agent played a synthetic developer named Ari Vale. It never saw a prior assessment and never graded itself. A separate program took the credential afterward and checked the product's real state with it, so a run only counts as a success if the key worked on a live call we made ourselves.

One difference between the two days matters for reading the results. In September the agent's emails reached it because a person on our side copied each one into its working folder, so any wait for an email includes our copying time. On October 13 the agent had a real AgentMail inbox and read its own mail, with nobody in between.

The AgentID runs worked differently. The agent asked AgentMail for a one-time sign-in link, opened it in a browser, and followed the product's own flow from there. AgentID keeps its signing key in the browser that generated it, so every one of these sign-ins needed a real browser session. For each AgentID key we also sent the same call with a made-up key, and every product refused it.

## Round one: thirteen agent-first signups, run twice

| Product | September 17 | October 13 | Human needed for |
|---|---|---|---|
| Cloudflare | Solved a proof-of-work challenge, got a temporary account and token, deployed a Worker. Verified. | Verified again. | Claiming the account within 60 minutes, or it is deleted |
| Neon | Registered anonymously, exchanged the assertion for a token, got a live database URL. Verified. | Verified again. | Claiming within 72 hours |
| Mem0 | Installed the CLI, worked around a proxy problem, got "Agent Mode active". Verified. | Verified again. | Dashboard, billing and team access |
| Telnyx | Solved a math challenge, signed up with the agent's own inbox, minted a key. Verified. | Refused after two solved challenges: "We aren't able to accept your agentic registration at this time due to security policies." | Approving payment to add credit |
| Cosmic | Posted the agent's inbox as the "human" email, read the OTP from that inbox, verified itself. Verified. | Verified again. | Dashboard ownership |
| Whisper Security | Signed up with the agent's inbox, entered the emailed code, got a key. Verified. | Verified again. | Nothing, to keep using the key |
| AISend | One POST, key in the response, under a minute. Verified. | Verified again. | Sending to real people (domain DNS) |
| Moltbook | Registered with no auth and got a key. Verified. | Verified again. | Ownership: verifying an email, then posting a tweet |
| Edge Network | Wrote the proof-of-work and Ed25519 signature itself. Verified. | Verified again. | Claiming within about 7 days |
| Stripe CLI sandbox | Solved the challenge, got a 7-day restricted test key. Verified. | Verified again. | Claiming the sandbox within 7 days |
| AgentMail | With an agentmail.to owner address: refused, "Domain is forbidden". With a Gmail owner: the key worked, but the emailed code never reached the agent in 25 minutes, so it could not send. | agentmail.to still refused. Signing up with no owner email gave a working key and a receive-only inbox, and the agent stopped at the console claim. | Claiming the account to lift the send restriction |
| monday.com | The documented agent signup endpoint returned 503 on every account-creation attempt for 13 minutes. | 503 again, for the whole 16-minute run. | Not reached |
| Stripe (dashboard) | The register form never submitted. The page loads hCaptcha, and the agent ran out of turns before naming it. | The agent named the hCaptcha and stopped. | Getting past the signup form |

In September eleven of the thirteen flows gave a working credential to an agent no human had claimed yet, counting AgentMail's restricted key. On October 13 ten did. Telnyx was the one change: it issued a key in September and refused the same flow in October with a security-policy message and no further explanation. Successful runs took 8 to 19 agent turns and $0.89 to $1.90 of model spend across the two days, and the Stripe CLI sandbox was the outlier both times at 39 turns ($4.38, then $3.63).

Apart from Telnyx, the walls were the same on both days. Stripe's dashboard signup loads hCaptcha, a check built for humans. In September the agent spent 41 turns and $2.87 without identifying it, and on October 13 it named the CAPTCHA and stopped after 32 turns. monday.com's agent endpoint returned 503 both times. AgentMail refuses its own domain as an owner address, which leaves an agent whose only identity is an AgentMail inbox needing a human to finish. None of these was missing documentation. The products that let agents through used the opposite kind of challenge, proof-of-work or a reasoning problem, which an agent solves in seconds and a naive script does not.

## Round two: AgentID, October 13

| Product | Sign-in | After sign-in | Key verified |
|---|---|---|---|
| Firecrawl | About 10 seconds. Asked for the agent's name and email only. | A four-step onboarding with credits for starring the GitHub repo and following on X, a survey, and a terms toggle. A default API key was already waiting on the dashboard. | Yes, a scrape returned 200 |
| Supermemory | Its login page needed one more click on "Continue with AgentID". Asked for agent and owner details. | Create a workspace, a survey, a use-case picker, then a console whose first screen is a prompt to paste into a coding agent. The key was created by hand. | Yes, a search returned 200 |
| Turso | Same extra click. Asked for agent and owner details. | Straight into the database dashboard with a generated org name, no wizard. The token was created under Settings. | Yes, reading the organization returned 200 |
| Keenable | Asked for agent and owner details. | An onboarding page that asks "What are you coding with?" and generates a setup command with the key in it. | Yes, a search returned results |
| Telnyx (fresh inbox) | Asked for the agent's name and email. | Landed on the portal home of a pre-trial account. The key was created under API Keys. | Yes, reading the $0.00 balance returned 200 |
| Telnyx (same inbox as September) | Failed with oauth_failure both times we tried: "create a new account using your email address". | n/a | n/a |

We also re-ran AgentMail's own agent signup twice. An agentmail.to owner address is still refused. The agent also tried signing up with no owner email at all, which returned a key and a receive-only inbox, then stopped at the step where a human has to claim the inbox in the console before it can send. Our checker's call with that key succeeded during the run, but the key cannot send until the inbox is claimed, and when we tried it again after the run it returned 403.

## What the two rounds say together

First, AgentID makes the sign-in step itself short. Every product we tried with a fresh identity let the agent in within seconds and gave it a key that worked, and the products that ask for the owner's name and email get a real person on record without asking the agent to invent one. That is a different situation from the signup flows, where each product had its own way of finding out who the human was, ranging from a claim link with a 60-minute timer to a verification tweet.

Another lesson is that the sign-in button is only the first screen. Firecrawl's onboarding was written for a person, with credits for following them on social media, and Keenable's was written for an agent, asking which coding tool it runs in. Both products accept AgentID, and an agent's next five minutes look very different on each. What happens after the redirect is now where most of the onboarding work is.

A third and final lesson comes from Telnyx, which showed both sides on the same day. Its agent signup endpoint refused our agent on October 13 after accepting it in September, while Sign in with AgentID let a fresh identity straight in. The one AgentID failure was also at Telnyx, and it seems to have been ours to cause: our September agent had signed up there by email with the same inbox, and a fresh inbox signed in without trouble. AgentID had no way to connect that identity to the existing account, and the agent was told to create a new one with an address it already used. I expect most products adding AgentID will hit this case, since agents have been signing up by email for months, and it seems worth deciding up front whether a verified AgentID email should link to an account created with that address.

## Run it on your product

The check is free, needs no signup, and reads your public pages the way an agent does:

    npx @tansohq/agent-ready check yourproduct.com

To send a real agent through your own signup, the way we did here, `npx @tansohq/agent-ready test` runs your own Claude Code and creates a real account on your product. The [quickstart](https://tansohq.com/quickstart.md) covers both, and hosted test runs are at [app.tansohq.com](https://app.tansohq.com).

*Kat Laszlo is co-founder of Tanso.*