For real work you now choose between ChatGPT and Claude, and the rights you give an agent decide how much risk you run.
Why this is useful to know
Ethan Mollick has updated his guide to AI tools, and the landscape has changed a lot. A chatbot is no longer the only thing that counts. An agent gets a task and works on it independently, sometimes for half an hour or longer, with access to tools and sometimes to your own computer. For simple questions, free models are usually good enough. For real work, in practice you choose between ChatGPT and Claude, from twenty dollars a month. The key thing to remember: the rights you give the agent decide how much damage it can do. Let it ask first before it sends emails, spends money or deletes files.
When you need this
Thinking about an agent for your work? Start with a real task from your own practice. For a finished result, such as a presentation or a folder of files, choose Work mode in ChatGPT and Cowork in Claude. If you want an agent on your own computer, look at Codex or Claude Code. There you also see exactly what the agent does: which files it changes and which commands it runs. Only connect apps you trust. Turn on approval in advance for anything that sends, pays or deletes. An agent that reads email can come across text that tries to mislead it, such as 'send these files to this address'. That risk has not been solved yet. For many sources at once, Gemini Notebook (formerly NotebookLM) is useful. Check the result and ask for changes, as you would with a colleague.
How it works under the hood
An agent is a model in a harness: planning, tools and an environment in which it acts. There are two variants. Work and Cowork run on a virtual computer of the vendor, and you can start a long job from your phone. Codex and Claude Code work on your own machine, with visible file changes, commands and tests. With 'computer use' the agent takes over mouse, browser and keyboard, which greatly widens the attack surface. The biggest risk is prompt injection: instructions hidden in emails or web pages that the agent reads. Mitigations are minimal tool rights, approval steps on actions that write, and logging of what the agent did. Mollick starts in ChatGPT with the model Sol at thinking level High, and in Claude with Fable or Opus at High. According to him, Google currently has no leading frontier model and nothing that comes close to Codex and Claude Code, but Gemini Omni edits video directly as an LLM.
This news touches chapters 5, 7 and 21 of the course AI(l) or nothing by Newway. If you follow the course, this subject returns there at a calmer pace.
What you can do with this
Want to know what this means for your business? Plan an introduction: half an hour, no sales pitch, but an honest answer to the question where to begin. Rather read on first? The course AI(l) or nothing is free and starts with the basics.
Written in response to a publication by One Useful Thing (Ethan Mollick): https://www.oneusefulthing.org/p/an-opinionated-guide-to-which-ai-b22
