Where the workflow shifted
Anthropic's computer-use docs describe an application loop that receives tool requests, executes screen or input actions, captures results, and returns them to the model.
Browser-agent reliability comes from observe, act, return evidence, and observe again, not from generating a long click script once.
Tool names are not outcomes
The signal matters when it changes how a team ships, reviews, or recovers work, not when it only names another tool.
Check permissions and failure
- Record each action, screenshot, page state, error, and next-step reason during browser tasks
- Keep the test narrow: one low-risk task or tool entry before connecting permissions, logs, failure handling, and human takeover to production
What still needs proof
If only the final result is stored, failures cannot be traced to model judgment, page drift, or environment errors. Keep the original source open so the announcement, the evidence, and this site's interpretation stay separate.