Spawned by a Prompt: OpenAI’s Agent Mode Feels Like a Mini OS

By Bharat Sharma 4:37 am AEST

OpenAI just dropped something quietly radical — a new Agent Mode inside ChatGPT that lets you spawn a task-specific virtual assistant that actually does things for you. This isn’t just upgraded autocomplete. This is prompt-to-action computing. And it feels like the start of something much bigger.

In their recent 30-minute YouTube demo, OpenAI showed off an assistant that can reason, browse, click through websites, summarize long documents, edit Notion pages, build Google Slides, update calendars, and even write and execute live code — all based on a single prompt. The Agent pauses for your permission before anything sensitive, walks you through its logic, and visibly performs each step like a fully-formed little OS running inside the chat.

What blew me away most wasn’t just the capabilities — it was the UX. It feels deliberate. The UI narrates each action, doesn’t move ahead without you, and makes even complex tasks like multi-tab research or calendar integration feel completely frictionless. You describe what you want — “Summarise these job descriptions and draft a personalized email to each recruiter” — and it just gets to work.

It’s funny timing too — just a month ago, Andrej Karpathy gave a talk at Y Combinator’s Startup School where he drew a direct analogy between large language models and operating systems: calling them abstract machines that run software described in natural language. And now here we are, watching OpenAI build exactly that: a task-running, sandboxed agent with memory, reasoning, execution, and security layers. It’s not just a metaphor anymore.

Agent Mode doesn’t get everything right yet. There are occasional UI fumbles, logic quirks, and performance gaps. But this isn’t a finished product — it’s an early glimpse into what AI-powered software execution might look like. And it’s already working better than most interns I’ve met.

The operating system of the future might not be a grid of icons.

It might be a blank prompt and a smart agent that listens, reasons, and acts.

And I can’t wait to use it and make an updated post based on actual experience.