Fusion harness

Fusion harness is the task-processing framework built by Cognition for its coding agent Devin. To use a cooking analogy, it's not about the chef's (the AI model's) skill itself, but more like a kitchen manual that decides when ingredients come out, in what order they're prepped, and when the stove gets turned on. Even with the same chef, a well-designed manual lets you make the same dish more cheaply, without wasting time or ingredients — and that's essentially what a fusion harness does.

When a coding agent takes on a task, it internally calls the model dozens or even hundreds of times, repeating a cycle of writing code, running it, and fixing errors. If the procedure that decides how efficiently and how often to call the model is poorly designed, costs can balloon even when producing the same result. Cognition refined this very procedure and stated that it achieved similar performance at a lower cost than simply using Anthropic's latest model directly.

In other words, a fusion harness isn't a new AI model — it's an operating method that determines how economically and cleverly an existing model is put to work. Separate from the competition over model performance itself, it's an example showing that carefully designing this kind of execution framework alone can cut costs significantly.

As one article put it: "On the same FrontierCode 1.1 Extended benchmark, the fusion harness scored 63.2 at $1.43 per task." A common misunderstanding here is thinking the fusion harness is itself a new AI model. In reality, it uses Anthropic's Fable 5.1 model as-is — Cognition only optimized, on its own, the procedure for calling that model and processing tasks.