What to Do When Your AI Agent Hits a Login Wall

You've built an agent that does real work, until it reaches a login screen, a two-factor prompt, or a CAPTCHA. Suddenly the whole run stops. This is one of the most common ways browser and desktop agents fail, and trying to brute-force past it is the wrong instinct.

Don't try to automate past it

OTPs, behavioral CAPTCHAs, security questions, and document uploads were built to require a real human. Automating around them is fragile, and modern sites are increasingly good at detecting it. Every workaround you build is one site update away from breaking, and one detection away from a blocked account.

The pattern that works: pause, hand off, resume

Instead of fighting the wall, acknowledge it needs a person and design for a clean handoff:

1. The agent detects the wall and pauses instead of failing the run.
2. It hands off to a human, sending a link to a live session, not a screenshot.
3. The person does the one human step: the login, the 2FA, the CAPTCHA.
4. The agent resumes with its session preserved, no restart, no lost state.

This "structured handoff" turns a total failure into a two-second interruption. The session stays alive, so the agent picks up exactly where it left off.

Make the handoff teach the fleet

Most human-in-the-loop tools stop there, every login wall pages a person, forever. The better version captures what the human did and turns repeatable parts into a reusable skill. A genuine authentication step (an OTP) will always need a person, and that's fine. But the navigation around it, dismissing the banner, finding the right button, the exact checkout path after login, only needs to be solved once. After that, the fleet handles it without paging anyone.

Security matters here more than anywhere

A fix session is exactly where a password gets typed. Any tool you use for this must never record typed text, only the shape of actions (a click, a keypress), never the characters. And fix targets should be restricted to legitimate pages; a handoff link that can be pointed at an internal address is a security hole.