Arbor: Tree Search as a Cognition Layer for Autonomous Agents

A new multi-agent framework called Arbor uses structured tree search as a shared cognition layer for autonomous agents, enabling coordinated optimization across large, stateful action spaces, according to a paper posted to arXiv on June 10 [1][2]. The framework departs from prior autonomous optimization systems that evaluate isolated targets without persistent state [2]. Arbor maintains an explicit search tree of scored hypotheses that functions as shared working memory across agents, evolving with each measurement and treating failures as diagnostic signals that reshape subsequent exploration [2]. The tree expands as prior successes shift the bottleneck distribution, allowing the system to adapt its search strategy over time [2]. Validation was conducted on full-stack LLM inference optimization, a domain that has historically required coordinated effort from engineering teams spanning the application, framework, compiler, kernel, and hardware layers [2]. Large language models are machine learning models with many parameters, trained with self-supervised learning on vast amounts of text [8]. The inference optimization challenge involves tuning performance across this entire stack, a task Arbor automates through its multi-agent design [2]. Arbor pairs two distinct agents in a checks-and-balances architecture. An Orchestrator agent drives optimization by delegating tasks to Domain Specialists across the inference stack, while a Critic agent safeguards stability through root-cause analysis, introspection, and measurement validation [2]. Neither agent can unilaterally drive the system [2]. Agent capabilities are decomposed into hard skills, representing domain expertise, and soft skills, which are coordination protocols that determine how contributions compose, enabling fully autonomous multi-day campaigns [2]. The framework achieved up to a 193% inference throughput-latency Pareto improvement over vendor-optimized baselines [2]. In contrast, a single agent operating without the Arbor harness plateaued at a 33% throughput improvement and crashed irrecoverably within hours [2]. The paper also reports that Arbor generalizes to multiple generations of hardware platforms, with run-to-run variance within 2 percentage points, indicating the method is hardware-agnostic and reproducible [2]. The paper appears on arXiv, a preprint repository that has integrated with platforms such as Hugging Face to make research more accessible through linked demos and community discussion [4][5]. Hugging Face Spaces allows researchers to build interactive applications that let users explore model outputs without writing code, and these demos can be embedded directly alongside papers on arXiv abstract pages [5][6]. The Arbor paper is indexed under the Computer Science and Artificial Intelligence categories on arXiv [1].