DataRobot, a leader in enterprise AI, has launched Syftr, a first-of-its-kind open source framework that programmatically discovers and optimizes agentic AI workflows for real-world applications. Aimed at AI practitioners and developers, Syftr makes it possible to identify the most effective combinations of components, tools, parameters, and strategies to achieve optimal performance in cost, speed, and accuracy.
Solving the Complexity of Agentic Workflows
With the explosion of agentic AI and retrieval-augmented generation (RAG) use cases, organizations are grappling with the challenge of managing and evaluating a vast and growing number of agentic system configurations. Syftr addresses this challenge with:
- Multi-objective optimization to balance trade-offs between task accuracy, latency, and compute cost
- A scalable, data-driven approach that simulates, tests, and prunes workflows using enterprise-grade data
- Compatibility with latest models and tools, ensuring current technologies can be tested in context
In rigorous industry-standard RAG benchmarks, Syftr identified configurations that reduced costs by up to 13x with only marginal compromises on accuracy—validating its value in enterprise deployments.
Features and Capabilities of Syftr
1. Discover Optimal Agent Workflows
- Multi-objective search using Pareto efficiency to identify the best configurations
- Tailored for use cases requiring accurate, fast, and cost-effective AI workflows
2. Reduce Cost and Time with Smart Pruning
- Bayesian optimization with Pareto pruner reduces non-promising trials early
- Achieves 80% reduction in compute cost and time during evaluation
3. Stay Ahead with Component-Agnostic Architecture
- Compatible with any embedding model, flow, or LLM
- Open to community contributions to keep up with cutting-edge tools and frameworks
4. Production-Ready Code Generation
- Generates clean, deployable LlamaIndex-based agent pipeline code
- Enables developers to move from experiment to production with minimal effort
Leadership Insight: The Need for Syftr in Enterprise AI
“Developers are navigating a space with 10²³ possible agentic architecture combinations. Syftr cuts through that noise,” said Venky Veeraraghavan, Chief Product Officer, DataRobot.
Veeraraghavan emphasized that Syftr is a paradigm shift—moving from manual experimentation to intelligent, scalable workflow discovery for production environments. This empowers developers to deploy effective agentic systems confidently and efficiently.
“Models are rarely used in isolation. Syftr enables end-to-end evaluation of workflows at scale,” added Debadeepta Dey, Distinguished Researcher at DataRobot.
“RAG and agentic applications are becoming more complex, and Syftr’s use of Ray and Ray Tune provides a scalable and powerful solution,” said Robert Nishihara, co-founder of Anyscale.
Availability and Community Access
- License: Open source with a permissive license
- Resources: Includes benchmark datasets and a DataRobot training dataset, all free on Huggingface
- Research Publication: Technical report titled Syftr: Bayesian Search for Pareto-Optimal Generative AI accepted for International Conference on Automated Machine Learning 2025
Accelerating Agentic AI with Syftr
Syftr marks a pivotal moment in agentic AI development. By providing a programmatic, multi-objective, and community-powered framework, DataRobot is enabling enterprises to move beyond trial-and-error workflows. With Syftr, AI teams can confidently build, benchmark, and deploy sophisticated agent pipelines that meet the dynamic needs of modern AI-driven businesses.
