SAGE cuts agent tokens 59% via self-built skill library

The 59% token reduction is the real headline here — if RL-trained skill reuse generalizes beyond AppWorld, this has direct infrastructure cost implications for anyone running agentic pipelines at scale.

• –SAGE extends GRPO with "Sequential Rollout": skills generated for Task 1 are available in Task 2, and Task 2's outcome reward flows back to reinforce good skill generation — true cross-task credit assignment
• –The skill library uses four operations (generate, use, update, save), letting agents patch failing skills mid-rollout rather than starting from scratch
• –Beating ReAct + GPT-4o / o1 / Claude Sonnet by 3x on scenario completion with a fine-tuned open model is a strong result for the "train smaller, smarter" camp
• –SFT initialization used Claude 3.5 Sonnet as oracle to bootstrap quality trajectories — underscoring how frontier models are increasingly being used to distill capabilities into cheaper models
• –GitHub repo is public at amazon-science/SAGE; reproducibility is high given the AppWorld benchmark is open