I tried OpenAI's new Atlas browser but I still don't know what it's for (www.technologyreview.com)

OpenAI quietly launched Atlas, a web browser that embeds ChatGPT and an agent meant to browse, answer questions, and perform tasks on the user’s behalf. In hands‑on testing the agent could, in theory, add items to a shopping cart, draft social posts, and summarize articles, but in practice performed slowly, produced irrelevant or repetitive results (e.g., re‑adding already purchased items or generating an off‑base Facebook status), and sometimes referenced the wrong page when asked to summarize content. The built‑in ChatGPT often felt no better — and sometimes worse — than using chat.openai.com, while the agent’s automation raised real UX and safety concerns (ten‑minute runs, messy carts, and potential for accidental actions). For the AI/ML community this rollout is notable less for novel model research than for its production‑level lessons: it highlights persistent problems in agentic UI design — grounding to live pages, session/context management, latency, relevance and hallucination control — and shows how embedding LLMs into client software surface privacy and telemetry questions. Atlas signals OpenAI’s push to capture richer behavioral data and integrate models into core apps, but it also underscores the engineering and safety work still needed before agents can reliably act on users’ behalf or justify switching from established browsers.