Arbor: Tree Search as a Cognition Layer for Autonomous Agents

Arbor is a new multi-agent framework designed to enhance autonomous agents' performance in large, stateful action spaces by implementing a structured tree search as a cognition layer. It features an Orchestrator agent for optimization and a Critic agent for stability, achieving up to 193% improvement in inference throughput-latency over vendor-optimized baselines while maintaining hardware-agnostic reproducibility with a run-to-run variance of just 2 percentage points. This framework allows for fully autonomous optimization campaigns, making it significant for practitioners aiming to improve LLM inference efficiency across various hardware platforms.