source: arxiv artificial intelligence: arbor: tree search as a cognition layer for autonomous agents
level: research
arbor is a multi-agent framework that uses tree search as a cognition layer for autonomous agents working in large, stateful action spaces. unlike prior systems that evaluate isolated targets without memory, arbor maintains an explicit search tree of scored hypotheses. this tree acts as shared working memory across agents, updating with every measurement. failures become diagnostic signals that guide future exploration, and successes shift the distribution of bottlenecks the system focuses on.
the framework was tested on full-stack llm inference optimization, a task that usually requires coordinated engineering across application, framework, compiler, kernel, and hardware layers. arbor uses an orchestrator agent to delegate tasks to domain specialists covering different parts of the inference stack. a critic agent evaluates the results, feeding back into the search tree. this setup allows the system to navigate a complex optimization space without human intervention.
by treating optimization as a search problem with persistent state, arbor can learn from past attempts and adapt its strategy. the tree structure lets agents explore multiple paths, backtrack from dead ends, and reallocate effort based on what has worked before. this approach moves beyond simple trial-and-error, making autonomous agents more effective in domains where actions have lasting effects and the best sequence of steps is not obvious.
why it matters: this method could make autonomous ai systems more reliable in complex, real-world tasks like system optimization, where remembering past attempts and learning from failures is crucial.
source: arxiv artificial intelligence: arbor: tree search as a cognition layer for autonomous agents