Arbor: Tree Search as a Cognition Layer for Autonomous Agents

Arbor introduces a multi-agent framework using tree search as a cognition layer for autonomous agents operating in complex action spaces. The system achieves 193% inference throughput-latency improvements over vendor baselines through coordinated Orchestrator and Critic agents, demonstrating reproducible, hardware-agnostic optimization across multiple hardware generations.

Arbor represents a meaningful advancement in autonomous agent coordination for computationally complex optimization tasks. Rather than treating optimization as isolated, stateless problems, the framework maintains a shared search tree that evolves continuously, enabling agents to treat failures as diagnostic signals that guide future exploration. This architectural approach mirrors how human engineering teams coordinate across domains, but operationalizes it through formal agent protocols.

The technical significance lies in addressing a genuine bottleneck in LLM inference optimization. Historically, achieving peak performance required manual coordination across application, framework, compiler, kernel, and hardware layers—a time-intensive process requiring deep expertise. Arbor's decomposition into hard skills (domain expertise) and soft skills (coordination protocols) enables this multi-layer optimization to run autonomously over multi-day campaigns without human intervention.

The performance differential is striking: 193% improvement with the full multi-agent system versus 33% with a single agent that eventually crashes. This gap reveals the value of the Critic agent's checks-and-balances architecture—preventing unilateral decisions that destabilize the system. The reproducibility metrics (run-to-run variance within 2 percentage points) across hardware generations establish that the method generalizes beyond specific platforms.

For the broader AI infrastructure sector, this work suggests a path toward autonomous optimization of increasingly complex software stacks. As LLM inference becomes more performance-critical for applications, automated approaches that eliminate manual engineering cycles could significantly reduce operational friction. The framework's applicability extends beyond inference to other multi-stack optimization problems, potentially influencing how AI development teams approach infrastructure challenges.

• →Arbor's tree-search cognition layer enables autonomous multi-agent coordination across complex, stateful optimization domains
• →Achieves 193% Pareto improvement on LLM inference optimization compared to vendor baselines through coordinated Orchestrator and Critic agents
• →Critic agent prevents system crashes and instability that plague single-agent approaches, improving stability by 6x relative to baseline improvement
• →Hardware-agnostic design demonstrates reproducibility across hardware generations with variance under 2 percentage points
• →Framework generalizes beyond inference optimization to multi-layer software stack problems requiring coordinated autonomous decision-making