MIT and Microsoft's Murakkab auto-optimizes AI agent workflows for speed and efficiency

Agentic workflows—AI systems that chain multiple models and tools to complete complex tasks like video analysis—are growing in complexity and becoming the backbone of cloud applications. However, manually designing and deploying these workflows often leads to wasted computation, energy, and cost. Researchers from MIT and Microsoft introduced Murakkab, a system that streamlines this process. Developers describe their application intent in high-level terms, and Murakkab automatically determines the optimal combination of AI agents, models, tools, and hardware configurations. It also dynamically adjusts resource allocation at runtime based on user-defined priorities—such as minimizing cost or maximizing speed—without requiring developers to hard-code every detail.

When tested on several agentic workloads, Murakkab significantly reduced the number of computational units needed for deployment, cutting energy requirements and costs compared to traditional approaches while maintaining performance. The system addresses a key challenge: the vast configuration space of black-box models and diverse tools from different providers. If a new, more efficient model emerges, Murakkab can incorporate it without manual rework. Lead author Gohar Chaudhry (MIT EECS) and senior author Ricardo Bianchini (Microsoft Azure) presented the work at the USENIX Symposium on Operating Systems Design and Implementation. The research highlights a practical path to making increasingly complex AI systems more sustainable and cost-effective for cloud providers and enterprises.