New scaling law prioritizes feedback over model size

The "Scaling Laws for Agent Harnesses" paper marks a pivotal shift in agentic AI, moving the industry's focus from raw model power to the efficiency of the surrounding execution layer.

• –Effective Feedback Compute (EFC) solves the "noisy compute" problem by only measuring computation that yields informative, non-redundant feedback.
• –While traditional scaling metrics fail to predict success, EFC coordinates show a 0.97 correlation with task completion across different model families.
• –The research highlights a "Routing Law" where accuracy decays logarithmically as tools are added, suggesting a hard limit on monolithic agent capabilities.
• –CheetahClaws serves as an open-source reference for building auditable, modular harnesses that prioritize feedback quality over token volume.
• –This work provides the mathematical foundation for why systems like Claude Code or Cursor outperform basic chatbot-style agents.