ServiceNow's AutoSynthData turns agent failures into training data
Diagnose where an agent fails, then auto-build the tasks to fix it — plus live data in Amazon Quick apps, Shopify's chat-to-store Canvas, and local inference.
Copy markdown
Turn agent failures into a training set
ServiceNow's AutoSynthData diagnoses where your agent falls short against a stronger teacher model, then generates and verifies targeted tasks to close the gap. In tests it lifted a Gemma-4-26B agent's Pass@1 by 7.2 points — a 35% relative gain that closed 59% of the measured gap — from 2,000 synthetic samples built in 18 hours; an IT-service-management agent climbed from 18.8% to 27.2%. The EnterpriseOps-Gym dataset is live on Hugging Face now.
Amazon Quick apps now serve live, per-user data
Amazon Quick's no-code AI apps can now query governed datasets in real time instead of build-time snapshots, and every query runs as the person viewing — so each user's row- and column-level security decides exactly what they see. Describe the app in plain language and ship a live dashboard or deal tracker with no backend to wire up.
Shopify Canvas builds a storefront from a chat
Shopify's new Canvas lets you stand up and restyle a custom store by describing it, with the layout redrawing visually as you talk. It collapses the theme-and-code work that used to gate custom storefronts — a realistic weekend path from idea to a live, on-brand shop for a solo founder.
Magnitude tops HN: one command to run models locally
Magnitude, an Apache-2.0 local inference engine, hit the Hacker News front page: it profiles your hardware, picks and tunes the best open models for it, then serves them to agents like Claude Code, Cline, and Pi — no API keys, no token costs. A fast way to take an agent workflow fully offline on your own GPU.