New AI framework boosts performance over competitors

Researchers have developed Arbor, a novel AI optimization framework designed to enhance the efficiency and reliability of AI agents in complex tasks. Unlike traditional methods that rely on trial-and-error, Arbor employs a cumulative learning process, organizing hypotheses and experiments into a structured tree. This allows the system to learn from past failures and make informed improvements over time. In tests, Arbor demonstrated over 2.5 times the verifiable performance gains of leading AI coding agents like Codex and Claude Code, operating under identical compute budgets. The framework separates strategic direction from execution, using a coordinator agent and isolated executor agents to explore multiple hypotheses concurrently, ensuring clean attribution and preventing repeated mistakes.