Issue 2026-09-28 · Industry · 产品 · 研究

Weights & Biases demos self-improving research agent Arya

Weights & Biases demos self-improving research agent Arya

Zubin Aysola of Weights & Biases demoed Arya, its now generally available self-improving research agent, at AI Engineer. Arya logs production traces in Weave and rewrites itself against 886 YAML tasks, scoring around 66% on some task sets. Its core is a flywheel where production traces and offline simulation share identical agent code, kept in sync by a four-hour mirroring job.

StartupHub.ai34 d ago
Read original ↗