Radar — 6 Jul 2026

🛠️ Development

• When agents should reach for a tool instead of thinking harder
  New ICML paper finds a 19–31 step 'deterministic horizon' past which handing state-tracking to tools beats longer reasoning — tool-using agents hit 86–94% accuracy versus 24–42% for pure CoT.
• A 5-step blueprint for agents that learn code skills for life
  A widely-shared post walks a five-step loop — generate a task at the agent's level, write a code skill, run it, fix it from errors, save it to a library — compounding skills without weight updates.
• The 'best coding agent' may be the note format, not the model
  A popular thread on a 16-page paper: same model, same task, and changing only how the previous session wrote its notes cut token use 42–63% for the same output.