Executable attention is a term Naresh Jain uses in Are Harnesses Worth the Cost? for the effect of a behavioral harness on a coding agent: the harness turns written requirements into environmental feedback, so that when a sensor goes red a requirement the agent was given but did not fully implement is brought back into focus and the agent gets another opportunity to act on it. The term pairs with what the author calls an attention lapse — code that omits or incompletely implements a requirement that was explicit in the agent’s instructions. In his interpretation the harness serves as an external memory and attention aid.
Usage
The term comes from an experiment in which coding agents built the same API service with and without a behavioral harness of Guides and sensors. A cost-optimized model’s unharnessed runs missed details that the brief clearly specified — a deferred-creation timeout workflow, a Kafka wire contract, exact error responses — while the harnessed runs all passed the external evaluation. The author offers executable attention as his reading of why, and says that the experiment did not establish the cause of the lapses; he suggests that juggling library choices, boilerplate, dependency injection and asynchronous wiring across a growing context may contribute.
When It Applies
As the author frames it, executable attention is what sensors provide when they run during development, and it matters for models whose unaided attempts produce inconsistent outcomes. For frontier models that solve the task in one pass, he recommends running the same sensors once as a final verification gate instead, because running them continuously added overhead without a detected benefit. It assumes the checks encode the right expectations: they can enforce only what humans have written into the contracts and examples, which need reviewing and maintaining alongside the implementation. The idea rests on a single experiment — 10 runs per model on one greenfield Java service with an unusually detailed brief — and is presented as the author’s interpretation and a conditional hypothesis, not an established result.