Feature request: Interrupt and take over

[ArmanJR]

Firstly, great job on the library and agent!

What's really stopping me from using this is that you can't take over the agent's active working env. It's a problem especially when you want to log in to websites. Most websites where these automation tasks can shine in are protected with auth; and sending plain passwords over the initial prompt isn't really ideal.

[ArturMakly]

This. need a simple way to add additional validation code sent to my email. or whatever that will happen during the session.
you need a human in the middle with most sites these days especially for registration or ID auth.

[ElasticBottle]

interesting, @ArmanJR @ArturMakly can you guys share some pseudo code of how you envision doing it given the API today?

Hear your use case loud and clear. Want to get a sense of what the intuitive way of adding extra information mid way looks like to y'all and design around that

[ArmanJR]

Not sure how the code would look like, but chatgpt's agent is doing it now and I think that's good enough. Basically, you tell the agent to pause, it will hand over the mouse and keyboard to you, and when you're done it will continue. A more sophisticated feature would be the case where the agent automatically detects that it now needs a human action, pauses the job and sends a notification.
I believe scenarios where human action is needed are well-known already. Auth, OTP input, payment, etc are all steps where users don't feel comfortable sharing sensitive data with AI, especially non-local models.

[ArturMakly]

Peep these: https://docs.skyvern.com/credentials/introduction