AI Agents Break One Step After the Demo Ends
The Second Click Problem
Every AI agent demo looks impressive. The agent navigates to a page, fills out a form, clicks submit. The audience applauds. But in production, the agent fails on the confirmation dialog that appears after submit. Or the redirect. Or the email verification step. Or the CAPTCHA.
This is the second click problem - agents that work perfectly for the demo path but break the moment reality diverges from the script.
Why Demos Lie
Demo environments are controlled. The page loads instantly. The form has no validation errors. The network never drops. Pop-ups do not appear. Session tokens do not expire.
Production has none of these guarantees. Real workflows include:
- Cookie consent banners that block interaction
- Session timeouts mid-task
- Dynamic content that loads after the initial page render
- A/B test variants that change the UI
- Rate limiting after repeated requests
Each of these is individually simple. Together, they create a reliability gap between demo and production that most agents cannot bridge.
Closing the Gap
Test past the happy path. For every workflow, identify the three most common failure points after the primary action and test against all of them.
Build recovery into the agent. When a step fails, the agent should not restart from scratch. It should recognize where it is, assess what went wrong, and retry the failed step with a different approach.
Use deterministic checks between steps. After each action, verify the expected state before proceeding. Did the form actually submit? Did the page actually navigate? Is the confirmation element present?
Log every real-world failure. Each production failure is a test case for the next version. Build a failure catalog and test against it regularly.
The Standard
Your agent is not production-ready until it handles the second click - the step after the demo - as reliably as the first.
Fazm is an open source macOS AI agent. Open source on GitHub.