Researchers from MIT and Microsoft have developed a system called Murakkab that optimizes the design and deployment of agentic workflows (AI-powered software systems chaining multiple models and tools). By enabling developers to describe desired tasks in plain language rather than hard-coding technical specifications, the system automatically selects the most effective models and tools. It also determines the optimal sequence of operations and hardware configurations to meet user priorities such as cost, speed, or accuracy.

Previously, developers had to manually define agentic workflows, a process that is often inefficient due to the complex, fragmented nature of black-box models and diverse external tools. Manual configuration frequently leads to over-allocation of resources, wasting energy and capital. Murakkab addresses this by dynamically making configuration decisions during deployment, allowing cloud providers to adjust resource allocation on the fly to maximize efficiency while adhering to user constraints.

When tested on diverse agentic workloads including video question-answering and code generation, Murakkab achieved significant efficiency gains. The system met user requirements while utilizing only about 35 percent of the computational units compared to traditional methods. Furthermore, it consumed roughly 27 percent as much energy at less than 25 percent of the standard cost. In one specific case, the researchers reduced energy consumption by more than an order of magnitude while maintaining accuracy within 2 percent of the original baseline.

The research team, led by electrical engineering and computer science graduate student Gohar Chaudhry, emphasizes that Murakkab provides cloud providers with necessary visibility into complex workloads, enabling more effective resource sharing. The project is supported by the Semiconductor Research Corporation and the U.S. Defense Advanced Research Projects Agency. Future development will focus on scaling the system for more complex workflows and larger computing clusters to further reduce the environmental and economic footprint of large-scale AI applications.

Researchers from MIT and Microsoft have developed a system called Murakkab that optimizes the design and deployment of agentic workflows (AI-powered software systems chaining multiple models and tools). By enabling developers to describe desired tasks in plain language rather than hard-coding technical specifications, the system automatically selects the most effective models and tools. It also determines the optimal sequence of operations and hardware configurations to meet user priorities such as cost, speed, or accuracy.

Previously, developers had to manually define agentic workflows, a process that is often inefficient due to the complex, fragmented nature of black-box models and diverse external tools. Manual configuration frequently leads to over-allocation of resources, wasting energy and capital. Murakkab addresses this by dynamically making configuration decisions during deployment, allowing cloud providers to adjust resource allocation on the fly to maximize efficiency while adhering to user constraints.

When tested on diverse agentic workloads including video question-answering and code generation, Murakkab achieved significant efficiency gains. The system met user requirements while utilizing only about 35 percent of the computational units compared to traditional methods. Furthermore, it consumed roughly 27 percent as much energy at less than 25 percent of the standard cost. In one specific case, the researchers reduced energy consumption by more than an order of magnitude while maintaining accuracy within 2 percent of the original baseline.

The research team, led by electrical engineering and computer science graduate student Gohar Chaudhry, emphasizes that Murakkab provides cloud providers with necessary visibility into complex workloads, enabling more effective resource sharing. The project is supported by the Semiconductor Research Corporation and the U.S. Defense Advanced Research Projects Agency. Future development will focus on scaling the system for more complex workflows and larger computing clusters to further reduce the environmental and economic footprint of large-scale AI applications.