Researchers have developed GATS (Graph-Augmented Tree Search), a planning framework that achieves 100% success rates on complex AI agent tasks while requiring zero large language model calls during execution.

The system, detailed in a new arXiv paper by Maureese Williams and Dymitr Nowicki, addresses a key limitation of existing approaches like Language Agent Tree Search (LATS) and ReAct, which rely heavily on LLM inference during planning.

GATS combines systematic UCB1-based tree search with a three-layer world model. The first layer handles exact symbolic action matching, the second uses statistics learned from execution logs, and the third employs LLM-based prediction for unknown actions.

Performance benchmarks show dramatic improvements

On synthetic planning tasks with branching paths and dead-ends, GATS achieved 100% success compared to 92% for LATS and 64% for ReAct. The gap widened further during comprehensive stress testing across 12 challenging scenarios including coding workflows, web navigation, and long-horizon tasks.

GATS maintained its perfect success rate while LATS dropped to 88.9% and ReAct fell to 23.9%. The framework required zero LLM calls per task during planning, compared to 37 calls for LATS.

The deterministic nature of GATS produces plans with zero variance across runs, eliminating the stochastic behavior that plagues LLM-guided approaches. This consistency addresses a major reliability concern for deploying AI agents in production environments.

The research demonstrates that systematic search combined with learned world models can substantially outperform LLM-guided exploration for agent planning tasks. The approach offers a path toward more efficient and reliable AI agents without sacrificing performance.

The paper suggests this methodology could enable broader deployment of AI agents in scenarios where computational costs and reliability are critical constraints.