What's this about?
Ethan Mollick argues the new "work agent" products (OpenAI's ChatGPT Work, Anthropic's Claude Cowork) stumble because they're built as if non-engineers can't reason about tasks the way engineers do — a product-design critique rather than a claim about model capability.
What are "work agents" like ChatGPT Work / Claude Cowork?
These are agent products aimed at general knowledge work — not coding — that take on multi-step office tasks (research, drafts, spreadsheets, ops). Mollick's point: their UX borrows coder mental models (tasks, tools, plans) that don't match how most professionals think, so non-technical users bounce off.
Why does it matter?
It's the same thread as the "nobody uses agents" and "model-speak" debates: the bottleneck to adoption is increasingly design and communication, not raw intelligence.