The End of Harness Engineering? (Google's Harness²)

Cloud Codes · 9:16 · Watch on YouTube

Google Research’s Harness² edits the system around a frozen model—prompts, tools, memory, and related pieces—so one step lifts GLM-5 on Work Buddy bench from 64.2 to 75.7, past Gemini 3.5 Flash’s untouched 72.4, while...

Read the full summary on retube