From the archive
Friday, 20 November
Folded from 34 sources that day.
Sorted by quality and by how many sources covered it, the same for everyone.
-
AI is Changing the Cross-Platform Decision Matrix
An experiment rebuilding a .NET MAUI app with native UIKit and Android Views found agents made a second UI cheap, but maintaining it still cost time. A full rewrite took under four hours.
-
Anthropic says Claude writes and reviews 90% of its team’s code. The catch is how it checks the work.
Anthropic runs deterministic checks such as builds, linters and tests before model review, and has Claude exercise changes in running apps to verify agent-written code.
-
Introducing harness adapters: adapt your favorite model to your harness to improve performance and cost
RL-trained LoRA adapters per agent harness raise Qwen3.8-27B's Terminal-Bench 2.1 pass@1 under pi from 39% to 55%, while cutting average tokens per task from 2.9K to 1.5K. A reusable recipe is included.
-
Fragments: November 20
An experiment on 465 open-source pull requests found unhealthy code raised the cost of agent review by 22 to 35%. The roundup also covers AI for reading terms of service and slowing down AI software factories.
-
Scaling Black Friday pipelines with Managed Airflow agents
Google shows how agents in the Airflow console diagnose failed tasks and check DAGs for upgrades, while Data Agent Kit writes and deploys DAGs from an IDE, using a clickstream example.