Overview
Use the following decision sequence:
- Is there a documented request or local tool that performs the action directly? Use it when the
format is known and the task is stable.
- Does the task require interacting with a dynamic website, clicking controls, reading rendered
state, or navigating forms? Use browser automation.
- Is the task outside the browser or unavailable through a direct interface? Use computer
automation.
- Is the task recurring and high volume? After proving the flow with browser or computer
automation, inspect whether a stable direct request can replace the slower layer.
- Does the platform block or restrict automation? Treat this as a design and policy constraint
rather than an invitation to retry indefinitely. Repeated failing calls waste time and tokens.
For a new integration, begin by reading available documentation. If the documentation is not accessible through a simple page request because it is dynamically rendered, use a browser to retrieve it. This small research cost can prevent many failed implementation attempts.
For a browser prototype, record the exact inputs, states, and outputs. When the workflow is stable, inspect the underlying requests if the environment exposes them. Document the discovered interface inside the project so future runs do not repeat the same reverse-engineering work.