An opinionated guide to which AI to use to do stuff

The Trigger: A Guide’s Evolution Reflects an AI Watershed

Simon Willison noted that Ethan Mollick’s AI tool guide a year ago read like a chat app recommendation list—ChatGPT, Claude, Gemini, each paired with its strongest model. Today, that guide has turned into a manifesto for “agentic systems.” This shift marks the industry’s pivot from conversational Q&A to autonomous task execution. Simon’s commentary signals something important: users suddenly realize they’re not holding a chatbot anymore, but a “digital employee” that can operate a computer—yet the job titles given to this new employee by vendors are baffling.

Breakdown: How Do Agent Modes Actually Work, and Why Is the Naming a Guessing Game?

Ethan explains that to harness the “computer” capabilities provided by AI companies, you need to select specific modes in ChatGPT or Claude. In ChatGPT, that mode is called “Work”; in Claude, it’s “Cowork.” But that’s just step one. If you download the desktop app, you’ll encounter “Codex” (ChatGPT) and “Code” (Claude), and these desktop modes can directly control your computer, offering more power than their mobile counterparts. Simon highlights a wildly counterintuitive detail: switching from “Chat” to “Work” on the mobile ChatGPT actually lifts the internet access restriction on its Code Interpreter container; meanwhile, the desktop “Work” mode is really a friendlier skin over Codex, with much deeper capabilities. For regular users, this design is like an advanced exam question no one studied for.

Trend Insight: The Agentic Explosion and Lagging User Experience

This situation reveals a deeper trend: AI tools are evolving from “apps” into “operating systems.” Before, opening an AI chat window was like loading a webpage; now, the AI asks to take over your mouse and keyboard, read and write files, and search the web—essentially becoming a second user on your computer. But vendors have moved so fast commercially that they haven’t designed an intuitive interaction model for this “second user.” The naming confusion is just the tip of the iceberg. The deeper problem is that users don’t know what permissions they’ve granted, and the AI doesn’t know which mode it’s in. It harkens back to the early smartphone era when typical users were baffled by toggles like “mobile data” and “Wi-Fi.” Now, anyone trying to get work done with AI has become that early-adopter geek.

Practical Value: How to Dodge the Naming Traps and Actually Use AI Agents

If you want to dive into agent modes, Simon’s observations offer three practical tips. First, forget the names and remember the capabilities: mobile Work/Cowork essentially runs in a cloud sandbox, ideal for isolated tasks like document processing or data analysis; desktop Codex/Code can access your local environment, making it suitable for batch file operations or automation workflows. Second, start with mobile, because its permission boundaries are clearer and mistakes are less harmful; desktop modes are powerful, but a misstep could overwrite your files or send wrong commands. Third, pay more attention to vendor actions than names: when a mode silently lifts a restriction or gains a new ability, that’s the real version update.

Counterintuitive Angle: The Naming Chaos Might Be a Deliberate “Complexity Barrier”

Most people assume this mess is just a product manager’s failure, but AI companies might actually have an incentive to keep things complicated. Clear mode switching and permission management would require educating users about responsibility, which could hurt retention. Instead, offering a few vaguely labeled “mode” buttons and letting users learn through trial and error showcases the AI’s power more quickly. In other words, this confusion is an extension of the LLM “black box” applied to UX: you don’t need to understand why, you just click and watch it work. Understanding the roots of this naming chaos is really about understanding how the AI industry treats its users.

Ethan’s guide acts like a mirror, reflecting our collective excitement and bewilderment about AI. The next time you hesitate between ChatGPT Work and Claude Cowork, consider Simon’s wry observation: you’re not failing to tell the names apart; you’re witnessing the birth of a new era—and all newborns have names that are hard to remember.