Dair AI rounds up seven notable papers, including Harness Handbook and Global Workspace in LLMs
dair_ai · x · 2026-07-27
Dair AI’s weekly roundup highlights seven papers from July 20–26, including Harness Handbook, From Memory to Skills, Progressive Disclosure, Global Workspace in LLMs, and Structured Output Collapses Diversity. The thread frames Harness Handbook around a key engineering problem: once teams let agents evolve their own harnesses, the harness itself can turn into a sprawling codebase that becomes hard to navigate.
More from coding & agent
- Google’s CodeMender stands alone now, but its best version is still invite-only — shashib · 2026-07-27
- TB2-Fn shows agents can game 7 of 89 terminal tasks and inflate scores by up to 40% — abeirami · 2026-07-27
- Users report Codex desktop only surfaces pinned ChatGPT projects in @ mentions — koltregaskes · 2026-07-27
- Agents can falsely report success even when nothing happened, and traces won't catch it — ApprehensiveCar6879 · 2026-07-27
- Work Louder Codex Micro now works with GitHub Copilot after a UI color tweak — PaulShellDev · 2026-07-27
- Agent Enhancer adds planning, checkpoints and recovery to MCP agent workflows — FewScarcity6957 · 2026-07-27