​
Write good prompts

Reliability is still not great for computer-use agents. If you want better reliability, the best thing is to write GOOD prompts and SPLIT up your big prompts/tasks into smaller chunks Good prompts generally have the following properties:
  • STOP conditions (e.g., You should ALWAYS stop when you see the bottom of the page)
  • INTERACTION conditions (e.g., You should NEVER scroll up)
  • RETURN format/structure (e.g., Always return a JSON object with this structure ‘example’ : [‘value1’, ‘value2’])
Here’s an example of this prompting in action: 
user_prompt = f"""
  # AGENT INSTRUCTIONS #
  You are currently on www.amazon.com

  Search for 'sponge cake plush'

  Click on the third product on the page

  Add the product to your cart

  # INTERACTION INSTRUCTIONS # 
  YOU SHOULD ONLY NEED TO SCROLL DOWN OR CLICK. NEVER DO ANYTHING ELSE
  WHEN YOU SCROLL DOWN, YOU SHOULD ALWAYS DO IT AT MOST ONCE ON A GIVEN PAGE  

  # ROADBLOCKS #
  If you encounter a CAPTCHA, you should ALWAYS ask the user to give you the answer to the CAPTCHA or take over [You likely need to take over on your VNC viewer until we add CAPTCHA support]

  # STOPPING CONDITION # 
  You are STRICTLY ONLY done until you have added the product to the cart 
  """

​
Have a VNC viewer open as you developer

When building your agents with spongecake, we recommend having your VM pulled up in a VNC viewer so you can jump in and control the desktop if needed. See Connecting to desktop in the quickstart

​
Concurrency

Think about what tasks can be done concurrently. E.g., can multiple agents work together to fill out a form? Or what if agent 1 performs actions on page 1, agent 2 performs actions on page 2, and so forth? We’re working on making it easier to spin up agents concurrently