📡 Radar — 26 Aug 2026

The daily catch for 26 August — all days →

📡 26 Aug 2026

🛠️ AI Development

• Article

  SWE-bench Science: can coding agents fix real scientific software?

  A 119-task benchmark across 98 scientific repos where the best setup — Claude Code with Opus-5 (max) — still clears under 50% pass@1. The failure analysis, not the score, is the read.

• Post

  LangChain's playbook for turning production traces into agent evals

  Turning real traces and human feedback into synthetic environments and evals to measure and post-train agents — concrete harness-engineering craft, if vendor-flavoured (LangChain pitching its skill).

  via @Vtrivedy10 ↗