ChatGPT Work Can Browse, File-Mangle, And Multitask. Free Users Get A Taste. Paid Plans Get The Real Lab.
David Gewirtz, Senior Contributing Editor at ZDNET, outlines how ChatGPT Work functions as an agentic AI system capable of browsing the web, managing files, and executing multistep projects. Free users can try it with limited capacity while paid plans offer more substantial access. Gewirtz's guidance is blunt: start small, restrict permissions, and verify its output because the thing makes mistakes.
This story illustrates the principle of bounded autonomy in AI systems. The mechanism is called progressive trust deployment. You grant an agent narrow permissions, observe its behavior, and expand access only after it proves reliable. Giving an AI tool unrestricted access on day one is the computational equivalent of handing your car keys to a stranger who claims to be a good driver.
David Gewirtz at ZDNET published a practical guide on August 24, 2026 covering ChatGPT Work's agentic capabilities. He notes a surprising surge in non-programmers using tools previously dominated by coders, with both Anthropic and OpenAI responding to the trend.
- Open ChatGPT on the free tier and ask it to research a simple topic, like the history of the typewriter. Watch how it browses and synthesizes. This demonstrates agentic behavior at its most basic.
- Upload a small text file and ask ChatGPT to summarize it or extract specific information. Observe how it handles file management.
- Assign it a two-step task, such as finding a topic online and then writing a short summary based on what it found. Check its work carefully. This is your first multistep agentic project.