Radar — 7 Jul 2026

New Anthropic interpretability research describes a small, consciously-accessible slice of the model's processing separate from the rest — with implications for detecting deception.

An open study of building software with coding agents: eight sourced sub-reports on context, harnesses, loops and workflows plus a synthesis of open problems, traced to primary sources throughout.

A designer walks through a real landing-page workflow using Claude's design and code tools, arguing the job is shifting from producing pixels to directing and quality-checking the output.

A widely-shared take that defining what 'good' means is the hardest unsolved problem on most AI teams, and that PMs — not engineers alone — are best placed to own the evals that pin it down.