An Opinionated Guide to Which AI to Use to Do Stuff (Summer 2026 Edition)
Source
Ethan Mollick
Author
Ethan Mollick
Date
Key takeaways · AI-distilled
Work and Cowork run on the AI company's computers, so you can kick off a long job from your phone and close the app. Codex and Claude Code run on yours, which is what lets them work across many files over a longer stretch.
Work and Cowork hand you a finished artifact — a deck, an analysis, a folder of files. Codex and Claude Code expose the work itself: files changed, commands run, tests executed. Pick by whether you want to review the output or the process.
Permissions, not model choice, decide whether an AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → acts on your behalf. The author's ChatGPT actually sent mail to a colleague because he had granted send permission earlier; Claude, set to ask first, only left a draft.
prompt injectionAn attack that hides instructions in content an AI will read — a webpage, email, or document — tricking it into following the attacker instead of the user.Full definition → is the reason to keep approvals on. An agent that reads your inbox and browses the web will eventually hit text written by someone else trying to redirect it, so gate anything that sends, spends, or deletes.
Handed a finished book PDF, GPT-5.6 Sol in Codex worked 30 minutes, chased down 195 references, and returned notes with no invented page numbers — but so nitpicky the author rejected many. The job is managing, not chatting.
Terms in this piece · Glossary
prompt injection — An attack that hides instructions in content an AI will read — a webpage, email, or document — tricking it into following the attacker instead of the user.
open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
Why it matters
It gives practitioners an up-to-date, opinionated map of which agentic mode and thinking level to reach for on a given task, plus a candid risk framing for granting agents access to your machine and data — the kind of decision framework that's hard to assemble from vendor docs.
Key quotes
“Basically, an agentic system gives an AI a computer to use.”
“Download Blender and make an otter using a laptop on an airplane.”
“You can almost think of the AI agents as a team that you delegate work to.”
