Ethan Mollick: Claude Code is a preview of agentic work everywhere

Ethan Mollick's argument is that coding tools crossed a line, and the line is not the model. It is the harness around it: more autonomous work with self-correction, plus specialized tools the agent can reach for. His demonstration makes the point concrete. He asked Claude Code to build a passive-income business, then left it alone. It worked for 74 minutes, wrote hundreds of files, and deployed a working e-commerce site selling prompt packs with no human writing code.

He names the mechanisms that make an hour-long task hold together. Compacting lets the agent take notes and clear memory when context fills, like the character in Memento, so it does not lose the thread. Skills are reusable instruction sets the agent loads only when relevant. Subagents split work across separate instances. None of this is a smarter model; it is scaffolding that makes the same model finish.

Mollick's wider claim is that this pattern leaves coding. He quotes Andrej Karpathy on the profession being refactored so the human-written bits are sparse, and argues the next tools will wrap the same harness around other knowledge work. Read the full post on One Useful Thing.

Why it matters

If you are judging where agents are useful, stop benchmarking the model alone and look at the harness. The 74-minute unattended build is the kind of test that tells you more than a leaderboard, and it is one you can run on your own workflow.