Maestro: Joint Graph and Config Optimization for Reliable AI Agents (arxiv.org)

Researchers have introduced Maestro, a novel optimization framework designed to enhance the reliability of large language model (LLM) agents by jointly optimizing both the underlying module graph structure and the configuration of each node—such as models, prompts, and tools. Unlike prior approaches that focus solely on tuning configurations while keeping the graph fixed, Maestro addresses structural failure modes by simultaneously searching for optimal information flow and component settings. This holistic approach significantly improves agent quality under practical rollout and token budget constraints, making it a versatile, framework-agnostic tool for AI developers. Maestro leverages reflective textual feedback drawn from execution traces to guide its search, improving sample efficiency and enabling targeted fixes beyond numeric performance metrics. Benchmark testing on IFBench and HotpotQA shows that Maestro consistently outperforms leading prompt optimizers like MIPROv2 and GEPA variants by notable margins—up to 12% on average—while requiring fewer rollouts. Even when restricted to prompt-only tuning, Maestro maintains a competitive edge, underscoring the effectiveness of its joint optimization strategy. Demonstrations on real-world applications, including interviewer and retrieval-augmented generation (RAG) agents, highlight that integrated graph and configuration optimization can resolve deep structural issues unattainable by prompt optimization alone, marking a significant advancement for building robust, reliable AI agents.