Arbor framework uses tree search as cognition layer for autonomous AI agents

Arbor is a multi-agent framework that rethinks autonomous optimization by introducing structured tree search as a cognition layer. Instead of treating each optimization attempt in isolation, Arbor maintains an explicit search tree of scored hypotheses as shared working memory across agents. This tree evolves with every measurement—treating failures as diagnostic signals that reshape exploration and expanding as successes shift the bottleneck distribution. The architecture pairs an Orchestrator agent, which delegates tasks to Domain Specialists across the full inference stack (application, framework, compiler, kernel, hardware), with a Critic agent that safeguards stability through root-cause analysis and measurement validation. Neither agent can unilaterally drive the system, creating a robust checks-and-balances mechanism.

Validated on full-stack LLM inference optimization, Arbor achieved up to 193% inference throughput-latency Pareto improvement over vendor-optimized baselines—a dramatic leap from the +33% plateau of a single agent without the harness, which also crashed irrecoverably within hours. The framework generalizes across multiple hardware generations with run-to-run variance within 2 percentage points, proving hardware-agnostic and reproducible. By decomposing agent capabilities into hard skills (domain expertise) and soft skills (coordination protocols), Arbor enables fully autonomous multi-day campaigns that previously required coordinated teams of engineers. This work represents a significant advance in scaling AI-driven optimization to complex, real-world systems.