We Let OpenAI’s “Agent Mode” Surf the Web for Us—Here’s What Happened

Why It Matters

The feature demonstrates meaningful automation potential for end users while exposing practical constraints (session timeouts, occasional misclicks and ethical guardrails), signaling OpenAI’s push to make agentic AI a consumer product and raising questions about reliability and safety at scale.

Summary

OpenAI this week debuted Atlas, a ChatGPT‑integrated browser with a preview Agent Mode that can click, scroll and perform multi‑step web tasks for users. In hands‑on tests the agent completed varied jobs with mixed results—novice‑level game play (2048 score ~3,164; 7/10), strong playlist creation from live radio (captured songs across sessions; 9/10) and effective email scanning into Google Sheets before hitting session limits (12 rows from 164 emails; 8/10)—and it appropriately refused requests to vandalize wikis. The feature demonstrates meaningful automation potential for end users while exposing practical constraints (session timeouts, occasional misclicks and ethical guardrails), signaling OpenAI’s push to make agentic AI a consumer product and raising questions about reliability and safety at scale.