Researchers from the Massachusetts Institute of Technology and Microsoft have developed a new system that could make artificial intelligence agents faster, cheaper, and far more energy efficient.
The system, called Murakkab, is designed to optimize the way complex artificial intelligence workflows are built and deployed. These workflows, often called agentic workflows, are increasingly used to power advanced artificial intelligence applications that rely on multiple models, tools, and decision-making steps.
Agentic workflows are software systems in which several artificial intelligence agents work together to complete complicated tasks. For example, an application may need to analyze a video, extract important frames, generate a transcript, search for related information, and then answer questions about the video. Each of these steps may require a different model, tool, or computational resource.
As these systems become more advanced, they are also becoming more fragmented and difficult to manage. Developers often have to decide in advance which models to use, which tools to connect, what order the steps should follow, what hardware should run the system, and how much computing power should be assigned to each part of the workflow.
These choices can strongly affect speed, cost, and energy use. If too many resources are assigned, the system wastes computation, electricity, and money. If too few resources are assigned, the workflow may become slow or fail to meet the userโs requirements.
To address this problem, researchers created Murakkab as an intelligent system that can automatically design and optimize these multistep artificial intelligence workflows.
Instead of requiring a developer to manually specify every technical detail, Murakkab allows the developer to describe the goal of the application in plain language. The system then determines which models, tools, and hardware configurations are best suited for the task.
For instance, a developer building a video question-and-answer application might simply describe the need to extract key frames from a video, generate a transcript, and answer user questions. Murakkab can then identify the most suitable models and tools for each part of the workflow and decide which steps should run one after another and which can run at the same time.
This ability to run some components in parallel can improve performance and reduce delays.
Murakkab also adapts its decisions dynamically. If a new model or graphics processing unit accelerator becomes available, the developer does not need to redesign the workflow from scratch. The system can update its configuration choices over time and adjust to new technologies.
When a cloud provider deploys an application using Murakkab, the system can further optimize the workflow based on each userโs priorities. Some users may want the fastest possible response. Others may want to reduce cost. Some may care most about accuracy, while still requiring the system to respond within a certain time limit.
Murakkab can balance these trade-offs automatically. It identifies the best hardware allocation and deployment schedule in real time, creating a workflow that is ready for the cloud provider to execute efficiently.
According to Gohar Chaudhry, a graduate student in electrical engineering and computer science at the Massachusetts Institute of Technology and lead author of the study, agentic workflows are becoming very complicated and are quickly becoming central to the work of cloud providers.
He explained that energy use is now a major concern, especially because it is easy to over-allocate computing resources when deploying these workflows. That over-allocation leads to wasted energy and higher costs. A system that helps cloud providers use resources more intelligently, he said, benefits everyone involved.
The research team also included Adam Belay, an associate professor of electrical engineering and computer science and a member of the Massachusetts Institute of Technology Computer Science and Artificial Intelligence Laboratory; Ricardo Bianchini, technical fellow and corporate vice president at Microsoft Azure; and other researchers from Microsoft Azure.
Their paper will be presented at the United States Advanced Computing Systems Association Symposium on Operating Systems Design and Implementation.
The challenge Murakkab addresses is especially difficult because agentic workflows often depend on several black-box models and external tools. These models and tools may come from different companies, and each may have its own settings and configuration options.
Because the number of possible combinations is extremely large, it is unlikely that developers could manually find the most efficient configuration, even if they tried. A new artificial intelligence model might improve accuracy or reduce cost, but under traditional methods, adding that model would often require developers to rework the entire application.
Cloud data centers also face a visibility problem. When they deploy an application for a customer, they often cannot see enough detail inside the workflow to allocate computing resources in the most efficient way at the time of the request.
Murakkab was designed to solve both sides of this problem. It helps developers create workflows more easily and helps cloud providers deploy those workflows more efficiently.
In tests involving video question-and-answer tasks and code generation workflows, Murakkab met user requirements while using only about 35 percent of the computation required by other methods. It also consumed only about 27 percent as much energy and operated at less than 25 percent of the cost.
The researchers also found that Murakkab could help users make informed trade-offs. In one case, the system reduced the energy consumption of an agentic workflow by more than an order of magnitude, while causing only about a 2 percent drop in accuracy.
Murakkab was also able to discover an unexpectedly effective configuration for a model that selects video frames in a video question-and-answer task. According to the researchers, this kind of optimization would be extremely difficult for a developer to identify manually.
The team now plans to extend the system to more complex workflows and larger computing clusters. They also want to explore how Murakkab can be used to optimize new kinds of agentic applications.
The research highlights an increasingly important issue in artificial intelligence: as applications become more powerful, the infrastructure behind them must become more efficient. Artificial intelligence systems are not only judged by what they can do, but also by how much energy, time, and money they require to do it.
Murakkab offers a promising way forward by allowing artificial intelligence workflows to become more adaptive, resource-aware, and efficient at cloud scale.
In simple terms, the system helps artificial intelligence agents work smarter behind the scenes. By choosing the right models, tools, hardware, and execution strategy, Murakkab could reduce waste while keeping advanced artificial intelligence applications fast and reliable.