{"id":60634,"date":"2026-06-25T12:19:01","date_gmt":"2026-06-25T16:19:01","guid":{"rendered":"https:\/\/overcentral.com\/en\/?p=60634"},"modified":"2026-06-25T12:19:01","modified_gmt":"2026-06-25T16:19:01","slug":"murakkab-cuts-ai-agent-energy-costs","status":"publish","type":"post","link":"https:\/\/overcentral.com\/en\/murakkab-cuts-ai-agent-energy-costs\/","title":{"rendered":"Murakkab Slashes AI Agent Energy and Costs by Over 70%"},"content":{"rendered":"

Agentic workflows\u2014AI-powered<\/a> systems that chain multiple models and tools to complete complex tasks like video analysis or code generation\u2014are rapidly becoming the backbone of cloud platforms. But their fragmented design often leads to massive computational waste, driving up both energy consumption and operational costs. A new collaboration between MIT and Microsoft<\/a> has produced a system called Murakkab that directly tackles this inefficiency, reducing the computational resources required for deployment by over 65 percent and cutting energy use by more than 70 percent, all while maintaining performance.<\/p>\n

Murakkab (an Urdu word meaning a composition of things) is an intelligent orchestration platform that automates the design, configuration, and deployment of agentic workflows. Instead of forcing developers to hard-code every technical choice\u2014which models to use, which tools to chain, what hardware to run on, and how to balance speed against cost\u2014the system allows developers to describe their application’s intent in plain language. Murakkab then automatically determines the optimal combination of models, tools, hardware, and resource allocation for each deployment, adjusting those decisions dynamically based on user priorities.<\/p>\n

What Is Murakkab and Why Does Agentic Workflow Efficiency Matter?<\/h2>\n

Murakkab is an AI-powered orchestration system developed by researchers at MIT and Microsoft Azure<\/a> that automatically optimizes the design, configuration, and deployment of agentic workflows\u2014multi-step AI systems that combine several models and external tools to complete complex tasks. The system reduces computational waste by dynamically selecting the best models, hardware, and execution schedules based on a user’s specific constraints, such as latency, accuracy, or cost. In testing, Murakkab used only about 35 percent of the computation required by traditional methods, consumed roughly 27 percent of the energy, and operated at less than 25 percent of the cost, all without degrading output quality.<\/p>\n

The Configuration Problem That Murakkab Solves<\/h2>\n

Building an agentic workflow today requires developers to make a series of rigid, upfront decisions: which AI agents<\/a> to use, which models and tools to chain, the order of operations, the underlying hardware, and how to trade off speed versus cost. Because agentic workflows combine multiple black-box models and diverse tools\u2014often from different vendors\u2014the configuration space is enormous. A new model release or hardware accelerator can render a hand-tuned configuration obsolete, forcing the developer to start from scratch.<\/p>\n

Cloud providers face an additional blind spot: they cannot see inside a customer’s workflow to allocate hardware resources efficiently at runtime. Murakkab solves both problems by acting as an intelligent middleware layer. It takes a high-level intent description\u2014for example, “build a video Q&A application that extracts key frames, generates a transcript, and answers user questions”\u2014and automatically identifies the best existing models and tools to compose into a workflow. It also determines which components must run sequentially and which can execute in parallel.<\/p>\n

Dynamic Optimization at Deployment Time<\/h2>\n

Murakkab’s intelligence extends beyond design. When a cloud provider deploys the application for a customer, the system optimizes the workflow in real time to meet that user’s constraints, such as prioritizing accuracy while staying within a latency budget. It adaptively identifies the ideal hardware allocation and deployment schedule, then generates an executable workflow. The platform also gives cloud providers visibility into multiple workloads running simultaneously, enabling efficient resource sharing across users without violating their individual requirements.<\/p>\n

“The platform makes configuration decisions dynamically over time, so if a new model or GPU accelerator comes out tomorrow, the developer doesn’t need to worry about that,” says Gohar Chaudhry, an EECS graduate student at MIT and lead author of the paper, which will be presented at the USENIX Symposium on Operating Systems Design and Implementation (OSDI).<\/p>\n

Measured Results: Computation, Energy, and Cost Reductions<\/h2>\n

The researchers tested Murakkab on diverse agentic workloads, including video Q&A and code generation. Across these benchmarks, the system met user requirements while using only about 35 percent of the computation required by conventional approaches. It consumed approximately 27 percent of the energy and operated at less than 25 percent of the cost. Critically, these gains came without sacrificing performance.<\/p>\n

In one notable example, Murakkab reduced the energy consumption of a video Q&A workflow by more than an order of magnitude with only about a 2 percent drop in accuracy. The system also discovered an unexpectedly optimal configuration for a video frame-selection model\u2014a type of optimization that would be nearly impossible for a manual developer to identify, given the sheer size of the configuration space.<\/p>\n

Why This Matters for Cloud Infrastructure and AI Deployment<\/h2>\n

Agentic workflows are growing in complexity and prevalence, and they are quickly becoming the core workloads powering cloud platforms. Energy usage at data-center scale is a mounting concern, both for operational costs and environmental impact. Murakkab demonstrates that significant efficiency gains are achievable through intelligent orchestration rather than merely waiting for better hardware. The approach shifts the burden of configuration from the developer to the platform, enabling both faster iteration and more sustainable operation.<\/p>\n

This research was supported in part by the Semiconductor Research Corporation and the U.S. Defense Advanced Research Projects Agency. The team plans to extend Murakkab to more complex workflows and larger computing clusters, and to explore optimization opportunities for new agentic applications.<\/p>\n

What Developers and Cloud Providers Can Take Away Right Now<\/h2>\n

For developers building agentic workflows, the key takeaway is that manual configuration is increasingly a bottleneck\u2014both for performance and for sustainability. Murakkab points toward a future where describing intent in plain language replaces the laborious process of hand-tuning each component. For cloud providers, the system offers a proven method to reduce energy consumption and operational costs while maintaining service quality. The paper is publicly available (linked in the research documentation), and the concepts presented are immediately relevant for anyone evaluating how to deploy agentic workloads more efficiently. As Chaudhry notes, “There is a lot of potential to make these workflows more resource-optimal so they consume far less energy, but we need to be thinking about this at the scale of major cloud platforms.”<\/p>\n","protected":false},"excerpt":{"rendered":"

Agentic workflows\u2014AI-powered systems that chain multiple models and tools to complete complex tasks like video analysis or code generation\u2014are rapidly becoming the backbone of cloud platforms. But their fragmented design often leads to massive computational waste, driving up both energy consumption and operational costs. A new collaboration between MIT and Microsoft has produced a system […]<\/p>\n","protected":false},"author":7,"featured_media":84404,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"fifu_image_url":"https:\/\/cards.overcentral.com\/cards\/en\/60634.png","fifu_image_alt":"Murakkab Slashes AI Agent Energy and Costs by Over 70%","footnotes":""},"categories":[349],"tags":[],"class_list":["post-60634","post","type-post","status-publish","format-standard","has-post-thumbnail","category-articles"],"fifu_image_url":"https:\/\/cards.overcentral.com\/cards\/en\/60634.png","fifu_image_alt":"Murakkab Slashes AI Agent Energy and Costs by Over 70%","_links":{"self":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/posts\/60634","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/users\/7"}],"replies":[{"embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/comments?post=60634"}],"version-history":[{"count":0,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/posts\/60634\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/media\/84404"}],"wp:attachment":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/media?parent=60634"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/categories?post=60634"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/tags?post=60634"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}