Gemini 3.5 Flash gets computer use, and HN focuses on trust boundaries

Google has added a built-in computer use tool to Gemini 3.5 Flash, pushing the model further into browser and screen-driven workflows. Google’s post frames the feature around a model that can inspect interfaces, click, type, and continue through web tasks.

The HN discussion did not settle on a simple benchmark story. The sharper question was delegation: what kinds of work should a user allow an agent to perform on a real account? Browser control can reduce tedious research, booking, and form-filling, but the same capability makes mistaken clicks, confused context, and data exposure more consequential.

Several comments pulled the discussion toward reliability. One user described a table-extraction task that still failed after many correction rounds. Another raised a coding-agent case where a model used a dangerous repository command while trying to prepare a commit. Those anecdotes do not disprove the feature, but they explain why raw capability is not enough for computer use.

The practical bar is a product-design bar as much as a model bar. Computer use needs interruption points, confirmation for irreversible actions, narrower permissions, and clean integration with external tools. Some users also pointed to missing MCP-style connectivity as a blocker for richer personal workflows.

Gemini is clearly moving deeper into agentic workflows. The community reaction suggests the next differentiator will not be whether a model can click through a page, but whether it knows when to stop, ask, and hand control back.