OpenAI’s ChatGPT Agent Signals a New Era of Autonomous AI Work Tools

OpenAI just dropped its most ambitious tool yet. ChatGPT Agent launched Thursday, and it's not content to just answer questions anymore. This thing wants to control your entire computer.

The new agent can navigate your calendar, book restaurants, generate slide decks, and write research reports. It handles tasks while you work on other things. But here's the catch - it might take 15 to 30 minutes to complete tasks that would normally take you hours.

"Even if it takes 15 minutes, half an hour, it's quite a big speed-up compared to how long it would take you to do it," says Isa Fulford, research lead on ChatGPT Agent. Translation: grab a snack while your AI does the heavy lifting.

More Than Just Web Browsing

This isn't your typical chatbot upgrade. ChatGPT Agent combines OpenAI's previous tools - Operator and Deep Research - into something that can actually manipulate software. Where Operator could only click around websites, this new agent has access to terminals, APIs, and a full computer environment.

The model behind it has no fancy name. OpenAI just calls it "the model behind ChatGPT Agent." Marketing department must've been out sick that day.

During demos, OpenAI showed the agent planning date nights by cross-referencing Google Calendar with OpenTable. It generated research reports comparing Labubu toys to Beanie Babies. One executive started using it to request office parking every Thursday - apparently forgetting to do this manually was becoming a weekly problem.

Safety First (Because AI Controlling Computers Sounds Terrifying)

OpenAI isn't naive about the risks. The company activated safeguards typically reserved for models with "high biological and chemical capabilities." They don't think ChatGPT Agent can help novices build weapons, but they're not taking chances.

The agent asks permission before doing anything "irreversible" like sending emails or making bookings. Financial transactions are restricted "for now." There's also something called Watch Mode - if you navigate to financial sites, you can't leave the tab or the agent stops working. It's essentially a safety switch that kicks in when you visit financial sites.

Performance Jumps Are Real

ChatGPT Agent scored 41.6% on Humanity's Last Exam, a brutal test covering over 100 subjects. That's double what OpenAI's o3 and o4-mini managed. On FrontierMath, one of the hardest math benchmarks around, it hit 27.4% with tool access. The previous best score was just 6.3%.

These aren't perfect scores, but they're substantial improvements. The agent isn't replacing human intelligence - it's just getting better at specific tasks.

The AI Agent Arms Race

OpenAI isn't alone in this space. Anthropic released Computer Use last October. Google, Meta, and Amazon executives won't shut up about their agent strategies on earnings calls. Everyone wants to build the next JARVIS.

The trend started gaining steam after Klarna announced its AI agent handled two-thirds of customer service chats in one month - equivalent to 700 full-time workers. That got everyone's attention.

But here's the thing about AI agents: they've been promising the moon for years while delivering incremental improvements. Most still struggle with complex tasks and feel more like tech demos than useful products.

Rolling Out to the Masses

ChatGPT Agent is available now to Pro, Plus, and Team subscribers. Just select "agent mode" in the tools menu or type "/agent" to access it. Enterprise and Education users get it later this summer. Europe and Switzerland will have to wait - no timeline yet.

The tool works best when you can walk away from it. Kumar and Fulford emphasize this isn't meant for real-time interaction. Set it up, let it work, come back to results.

Still Has Growing Pains

Don't expect lightning speed. The agent prioritizes thoroughness over velocity. For online shopping, Fulford found it more thorough than using Operator alone, but that thoroughness comes with wait times.

The team of 20 to 35 people who built this thing focused on "optimizing for hard tasks" rather than quick responses. They're betting users will accept slower performance for better results.

Why this matters:

• We've had AI agents that could browse websites for a while now. ChatGPT Agent is the first one that can actually control your computer - clicking buttons, running programs, managing files. That's a big jump from what came before.
  
• The benchmark scores aren't just numbers. When an AI doubles its performance on complex tests, it usually means the underlying capabilities took a real step forward. We're seeing AI move from "sounds smart in conversation" to "can handle actual work tasks."