Trust Is Cheaper Than Control
Whenever I log in to claude.ai, I think about what I want to discuss and which model I want to choose. I look at the list of models and choose to stay away from Fable, after all it costs twice as much as Opus. When it comes to using a coding harness, where you have a lead model and a sidekick, I would have thought to go with Opus again over Fable. A little less performance at half the cost, right?
Wrong.
The engineers who make Devin tested this scenario. Their harness has a lead agent and a sidekick. They found that it costs less to run Fable as the lead than Opus. The Fable-led one performs 11% better than the Opus-led one and costs 9% less.
It is natural to think that maybe Fable simply delegates more than Opus. But the analysis of the results shows that both delegate the same amount. Rather, Fable delegates early and provides clear instructions along with constraints and a clear definition of done. Fable did not even make a single code edit about 81% of the time compared to Opus's 24%. Fable simply trusts the sidekick more and chooses not to micromanage.
We actually have real world examples of leaders with similar management styles. Reed Hastings led Netflix with the mantra "lead with context, not control". In his book No Rules Rules Reed describes a situation where one of his managers made what he considered an awful decision. The decision looked bad to him until he asked why, and it turned out reasonable given what the manager actually knew. The real problem was that senior leadership hadn't given him enough context to make a better call.
Hastings arrived at this philosophy over a career. Fable uses it effectively across thousands of sessions. When a management philosophy holds up that consistently, whether a person or a model is running it, that tells you something about the philosophy. Trust and clear context are not soft ideas. They are what works.
I think we will see more harness systems with this pattern. The frontier models decide what to build and what to constrain. The cheaper, specialised models do the building.
And maybe the lesson runs the other way too. If trust and clear context are what make Fable work, they are worth building our own workplaces around.
If you want a technical deep dive into the actual benchmark, click here.
