Agentic pipelines are becoming increasingly sophisticated, integrating components such as RAG modules (e.g., vector databases, embedding models, retrievers), as well as verifiers, rewriters, and rerankers. Each module introduces complex configuration and hyperparameter choices, resulting in a vast design space with different trade-offs between latency, accuracy, and cost. Syftr is a novel framework that automates this exploration of this space, using multi-objective Bayesian Optimization to surface Pareto-optimal agentic pipelines. In this presentation, we explore how pipelines running on the Cerebras Wafer-Scale Engine perform when optimized by Syftr for low-latency use cases.