Programmatically discover, optimize, and customize use case-specific agentic workflows
DataRobot, the agentic workforce platform, today announced syftr, a first-of-its-kind open source framework designed to identify performant agentic workflows for commercial use, now available. Syftr empowers AI practitioners to programmatically discover and implement the best combinations of components, parameters, tools, and strategies for agentic use cases, optimized for accuracy, processing speed, and cost.
As organizations increasingly explore agentic AI systems, practitioners and developers need to quickly evaluate the latest technologies and ensure that their agentic workflows are optimally performant for specific use cases based on model quality, cost, and desired behavior. Syftr addresses this challenge through a groundbreaking multi-objective approach that rapidly simulates possible configurations to identify the best AI workflows with enterprise data and optimizes for task accuracy, latency, and cost. In industry-standard RAG benchmarks, syftr identifies workflows that cut costs by up to 13x with only marginal accuracy trade-offs—delivering near-optimal performance at a fraction of the price.
“Practitioners and developers are navigating a constantly evolving AI ecosystem—on the order of 10²³ possible agentic architecture combinations—where the most obvious approaches often fall flat. Our mission is to cut through that noise and guide developers to the parameters that actually work for commercial use cases and production environments. With syftr, we’re changing that paradigm to make agentic AI useful, performant, and customizable for enterprises. For the first time, practitioners and developers can truly evaluate the full landscape of AI technologies against company data and implement use cases that balance accuracy, speed, and cost. Now with syftr, they can confidently and quickly implement agentic pipelines and take the guesswork out of manual experimentation," said Venky Veeraraghavan, Chief Product Officer at DataRobot.
Syftr streamlines the evaluation of entire agentic workflows through several key innovations. Now AI practitioners can:
Discover optimal agent pipeline patterns, components, and parameters:
- Multi-objective search: Leverage a novel approach using Pareto efficiency to rapidly generate and evaluate different workflow strategies, parameters, models, and components to find a configuration with optimal accuracy, cost, and latency.
Run computations efficiently with minimized costs:
- Bayesian optimization early stopping mechanism: Expedite search using the Pareto pruner technique to compare new subflows to a baseline benchmark, removing any new flows that do not show promise by meeting or exceeding the baseline. This process produces an 80% reduction in compute time and cost.
Evaluate and implement the newest techniques and technologies:
- Component agnostic: Evaluate any module, flow, embedding model, or LLM, ensuring even the most recent technologies are considered for optimization based on contributions from DataRobot engineers and the open source community.
- Agent pipeline code generator: Easily implement and finetune AI workflows by copying the generated production-ready LlamaIndex code.
“At today’s scale and pace of innovation, it’s impossible for developers to manually evaluate every new technique, tool, and LLM update. And while there are many benchmarks to evaluate model capabilities and performance in isolation, models are rarely used in a vacuum, particularly in the enterprise. Now, syftr is breaking down these barriers for the first time and enabling AI teams to explore large-scale workflow search spaces and deliver AI agents faster than ever before,” said Debadeepta Dey, Distinguished Researcher at DataRobot.
"RAG applications and agentic applications are exploding in complexity due to the number of moving parts and the number of decisions developers need to make. Syftr is an impressive framework that addresses the need to simultaneously optimize cost, accuracy, and latency in agentic applications. Its innovative approach relies heavily on Ray and Ray Tune to manage scalable search processes across CPUs and GPUs. I am thrilled that Ray is enabling such an innovative tool and I'm excited to see the AI community build on it," said Robert Nishihara, co-founder of Anyscale.
The syftr framework is available immediately as a permissively-licensed open source project with industry benchmark datasets and a DataRobot training dataset accessible for free on Huggingface. DataRobot also published a technical report syftr: Bayesian Search for Pareto-Optimal Generative AI that contains all the scientific details. The technical report has been peer reviewed and accepted for publication in the International Conference on Automated Machine Learning 2025.
The enterprise version of syftr will be available in fall 2025.
Learn more
To learn more about syftr, please visit the product page, review the report, and try today.
About DataRobot
DataRobot is agentic AI for the workforce. Our agentic platform enables frontline/business teams to develop, deliver, and govern AI agents and applications that work intelligently and securely with core business processes, infrastructure, and systems — maximizing impact and minimizing risk for organizations across industries. For more information, visit our website and connect with us on LinkedIn.
View source version on businesswire.com: https://www.businesswire.com/news/home/20250528824014/en/
