Eight ways to build a browser agent

There is no single right way to build a browser agent. Here are eight patterns we see in the wild, from the simplest to the most ambitious.

• Prompt-to-action: let the model emit clicks and types directly
• Tool-calling: expose navigation and extraction as functions
• DOM-to-JSON: hand the model a cleaned page and ask for data
• Vision-first: screenshot, reason, act
• Record-and-replay: capture a flow once, parameterize it
• Skill-based: call named, versioned tasks
• Human-in-the-loop: pause for approval on risky steps
• Fleet: run many of the above in parallel