Murakkab: the system that optimizes multi-step AI workflows

What happened?

A team of researchers from MIT and Microsoft has developed Murakkab, an intelligent system that automates the design and optimization of agentic workflows (workflows based on AI agents). These workflows chain multiple models and external tools to perform complex tasks, such as analyzing a video and answering questions about it. The system, presented at USENIX OSDI 2026, allows developers to describe in natural language what they want the workflow to do, without needing to specify technical details. Murakkab automatically selects the best models, tools, hardware configuration, and cloud resource allocation, adjusting in real time to user priorities (minimum cost or maximum speed).

The original MIT News article details that agentic workflows are becoming "extremely complex" and are quickly becoming the backbone of what cloud providers do. However, the way these highly fragmented systems are designed and implemented often leads to inefficiencies that result in wasted computation, energy, and costs. Murakkab addresses this problem directly, acting as an intelligent orchestrator that optimizes every aspect of the workflow.

Historical context and comparisons

Workflow optimization is not new: from business process management (BPM) systems to microservice orchestrators like Kubernetes, the industry has sought to automate resource allocation. However, Murakkab represents a qualitative leap by operating in the domain of AI agents, where the diversity of models (from GPT-4 to specific vision models) and tools (APIs, databases, search engines) multiplies complexity. Unlike previous approaches like AutoML, which optimize individual models, Murakkab optimizes entire chains of models and tools, considering both latency and energy cost.

Historically, resource overallocation has been a chronic problem in cloud computing. A 2020 study by the Uptime Institute found that 30% of servers were underutilized. With the advent of generative AI, energy consumption has skyrocketed: according to the International Energy Agency, training a model like GPT-3 consumed approximately 1,300 MWh, equivalent to the annual consumption of 130 US households. Murakkab attacks this problem in the inference phase, which represents the majority of operational consumption for cloud providers.

Why is it important?

Multi-agent workflows are becoming the backbone of cloud services. However, their fragmentation and complexity generate inefficiencies: developers often over-provision resources, wasting energy and money. Murakkab addresses this by reducing the number of computational units needed, resulting in significant energy and cost savings without sacrificing performance. According to Gohar Chaudhry, lead author of the study, "It's very easy to over-provision resources, wasting energy and money. Allowing a cloud provider to intelligently optimize these workflows is a win for everyone." At a time when AI energy consumption is a growing concern, this optimization is key to sustainability.

The potential impact is enormous. According to a 2024 Goldman Sachs report, data center electricity demand is expected to double by 2030, driven largely by AI. Murakkab could help mitigate this growth by reducing consumption per task. Additionally, by lowering operational costs, cloud providers could pass those savings on to customers, democratizing access to advanced AI.

How does it work?

Murakkab acts as an intelligent orchestrator. From a natural language description, the system explores the space of possible configurations—models, tools, hardware—and selects the optimal combination. Moreover, during execution, it dynamically adjusts resource allocation based on user priorities. This eliminates the need for developers to hardcode all technical decisions, making it easier to adapt to new models or changes in workloads. In tests with several real-world workflows, Murakkab reduced the number of required computational units by up to 50%, with a proportional decrease in energy consumption and costs, while maintaining the same accuracy and speed.

The paper presented at OSDI 2026 describes that Murakkab uses a reinforcement learning-based search approach to explore the configuration space. Unlike heuristic methods, this approach adapts to the specific characteristics of each workflow, achieving finer optimization. Experiments included typical workflows such as video analysis with questions, document processing, and multi-layer chatbots, consistently showing resource reductions of 30-50%.

Consequences for the industry

This advancement has direct implications for cloud providers, AI developers, and end users:

• For cloud providers (AWS, Azure, Google Cloud): it enables offering more efficient and eco-friendly AI services, reducing their carbon footprint and operational costs. Azure, as a co-author of the study, could integrate Murakkab into its platform, gaining a competitive edge over AWS and Google Cloud, which have not yet announced similar systems.
• For developers: it simplifies the design of complex workflows, accelerates time-to-market, and reduces the need for optimization experts. This is especially valuable for startups that lack resources to manually tune every configuration.
• For users: faster and cheaper applications with lower environmental impact. For example, a video analysis service could reduce its costs by 40%, making it affordable for small businesses.

Furthermore, Murakkab could pave the way for more sustainable AI, aligned with global energy efficiency goals. In a context where companies like Google and Microsoft have committed to being carbon negative by 2030, tools like this are essential to meet those objectives without sacrificing innovation.

What readers should know

Murakkab is not a commercial product, but a research prototype. However, the involvement of Microsoft Azure suggests it could be integrated into cloud services in the future. Developers should watch for similar tools that automate the optimization of multi-agent workflows, as they will become a standard for reducing costs and energy consumption. It is important to note that the system has not yet been tested in large-scale production environments, and its performance may vary under real conditions. Additionally, reliance on a centralized orchestrator introduces a single point of failure, although the authors claim the design is robust.

"Agentic workflows are getting very complicated and quickly becoming the backbone of what cloud providers are doing. Energy usage is a huge concern, so we need to be very careful about how efficient these workflows are." — Gohar Chaudhry, MIT

In summary, Murakkab demonstrates that it is possible to achieve efficiency without compromising performance, a critical balance for the future of AI in the cloud. As agentic workflows become ubiquitous, automated optimization will be a key differentiator for cloud providers. Next steps include testing in real environments and possibly integration with platforms like Azure Machine Learning. Developers would do well to familiarize themselves with these concepts, as the trend toward automation of optimization is unstoppable.