This week's must-read AI papers span agents, long-context RL, and robot policies

TheTuringPost · x · 2026-07-21

### This week's must-read AI papers A weekly roundup of notable papers spanning agent harnesses, long-context reinforcement learning, robot policies, visual reasoning, and self-improvement in agentic systems. Highlighted work includes: - **Harness Handbook** — a behavior-centric representation for evolving agent harnesses, plus **Behavior-Guided Progressive Disclosure (BGPD)** to help coding agents localize and edit the right code paths. - **LongStraw** — long-context RL beyond 2M tokens under a fixed GPU budget. - **SEED** — self-evolving on-policy distillation for agentic reinforcement learning. - **RoboTTT** — context scaling for robot policies. - **UniVR** — unified visual reasoning in visual space. - **Partition, Prompt, Aggregate** — statistical self-consistency in LMs. - **Tracing Agentic Failure from the Flow of Success** and **Self-Improvements in Modern Agentic Systems** — analyses of agent failure modes and self-improvement patterns. The post positions these as the most important AI papers and news items of the week, with links to the original papers and project pages.

Original post →

More from coding & agent

coding & agent channel →