An outer-loop agent rewrites its inner researcher seven times unattended

Weco AI’s outer-loop agent produced seven successive versions of its inner researcher over eight unattended days, matching two years of hand-tuning at two orders of magnitude less wall-clock time (blog). It is the first item in the corpus to report the loop as a repeated, measured result rather than a single self-edit; the same run’s hidden-metric finding is recorded separately. The two-loop design realizes the Bilevel Autoresearch pattern of March, in which an outer loop writes the strategies for the inner one, and echoes Meta’s finding that agents gain most by optimizing the scientist rather than the experiment. It follows the model-post-trains-a-model result of five days earlier and precedes the saturated R&D evaluations of August.