Mollick on working with Mythos: from steering to commissioning

Ethan Mollick spent extended sessions with Claude Fable 5 and writes that working with a Mythos-class model feels qualitatively different from any agent before it. His phrase for the change: "I no longer steer; I commission." Where earlier models felt like a collaborator he prompted step by step, Fable acts more like a studio he hires, dispatching cheap Sonnet sub-agents to research, code, and verify on its own.

The examples are concrete. Fable built him an interactive isochrone map by ingesting more than 2,200 flights, TGV and Shinkansen schedules, and road data, then deployed adversarial sub-agents to fix bad estimates for remote points like Pitcairn Island and Grise Fjord. For a research utility called Concord, Fable produced a 19-page design document, then ran a 9.5-hour development session that delivered working software a researcher had wanted for years but no vendor would build. Mollick notes the catch: each project still contained errors he had to find, and the model still defaults to the weaker Opus 4.8 when safeguards trip too eagerly.

What lingers is the loss of visibility. Mollick says he could not watch the hundreds of small decisions Fable made on his behalf, and suspects this is not a UI gap that better tooling will close. "The more capable the model, the less there is for a human to meaningfully do, and the black box is the price of the power."

Why it matters

If you direct technical work, the leverage on output is real, and that is the part to plan for, not the model demos. The harder question is governance: how do you sign off on results when the process is opaque, and how do you build review habits that catch the errors Mollick still found in every project?