📡 Radar — 10 Sep 2026

The daily catch for 10 September — all days →

📡 10 Sep 2026

🛠️ AI Development

News
AutoResearchExam: an open benchmark for AI research agents over 24-hour runs
An open benchmark runs agents on 29 ML research tasks over 24-hour windows, scoring generalisation on a hidden test set — models keep improving throughout, and step quality beats submission volume.
Article
Andrew Ng maps the concrete skills for using coding agents well
Five concrete skills for working with coding agents — directing the workflow, calibrating autonomy, reviewing output, customising the setup, and knowing how agents fail. Pointedly anti-hype.
News
Scanning the Harness: auditing defects in AI coding-agent configs
A study audits real defects in the instruction files, skills and harnesses that configure Claude Code and Cursor, and ships harness-eval, 'a linter for agent configs', with corpus and test scenarios.

Post
Beyond the Chatbox: designing AI agents for real actions, not just chat
A product designer walks through an AI-agent UX framework arguing not every interaction should be a chat — a prompt to design for uncertainty, approval and trust instead of another text box.
via @Manny_mikee ↗
Post
UX.md experiments: how much UX context changes what AI generates
A first-hand interactive report on how the amount of UX context you feed a model changes the product features it generates — small experiments, but concrete evidence behind the trending 'UX.md' idea.
via @AnthonyPAlicea ↗