Radar — 14 Jul 2026

🛠️ Development

• A 71-author field guide to code LLMs and coding agents
  A 71-author survey of the code-LLM lifecycle — data curation, pre-training, fine-tuning, RL and prompting — up to autonomous coding agents, focused on where benchmarks diverge from real deployment.
• Failure as a Process: an anatomy of CLI coding-agent trajectories
  1,794 hand-annotated agent runs on Terminal-Bench show failures usually start early but stay hidden until recovery is impossible — so reliability needs mid-run intervention, not end-state grading.