VisualWebArena agent

[gasse]

No description provided.

[ThibaultLSDC]

I have updated generic_agent to accept goal_images. it alters slightly the original behavior when using the screenshot as input.
original behavior:
System message
----text: system_prompt
Human message
---- text: prompt
---- image: screenshot

New behavior:
System message
---- text: system_prompt
Human message
---- text: prompt
---- text: IMAGE (1) - screenshot input # that would appear even without goal_image
---- image: screenshot
---- text: image (2) - image input # if there is a goal_image
---- image: goal_image 1
....