Researchers from MIT and Microsoft Azure have developed Murakkab, an intelligent optimization platform designed to streamline the deployment and resource allocation of AI agentic workflows, significantly reducing computational overhead and energy consumption. Agentic workflows, which chain multiple autonomous AI models and external tools to execute complex, multi-step tasks, are rapidly becoming foundational to cloud computing. However, their fragmented architecture traditionally forces developers to manually hard-code configurations, specifying hardware requirements, model selections, and execution sequences upfront. This static approach often results in severe resource overallocation, driving up costs and energy usage while struggling to adapt to new models or shifting user priorities. Murakkab addresses these inefficiencies by shifting configuration to a dynamic, intent-driven process. Developers outline their application objectives in plain language, and the system automatically selects the optimal suite of models, tools, and execution pathways. Crucially, the platform operates continuously during deployment, adjusting parallel and sequential processing logic, hardware allocation, and computational resource distribution in real time. By granting cloud providers visibility into multiple concurrent workloads, Murakkab balances competing constraints such as latency, accuracy, and cost without requiring developers to manage underlying technical details. Benchmarked across video question-answering and code generation workloads, Murakkab demonstrated substantial efficiency gains. The system achieved target performance metrics using approximately 35 percent of the computational units required by conventional methods, while consuming roughly 27 percent of the energy and operating at less than 25 percent of the traditional cost. In scenarios prioritizing sustainability, the platform reduced energy consumption by over an order of magnitude with only a marginal two percent decline in model accuracy. The architecture also automatically identified non-intuitive hardware and model combinations that would be impractical for manual configuration. The research, supported by the Semiconductor Research Corporation and the U.S. Defense Advanced Research Projects Agency, will be presented at the USENIX Symposium on Operating Systems Design and Implementation. Lead author Gohar Chaudhry, alongside MIT associate professor Adam Belay and Microsoft corporate vice president Ricardo Bianchini, emphasized the necessity of optimizing agentic systems at cloud scale as energy demands escalate. Future development will focus on extending Murakkab to support more complex workflows and larger distributed computing clusters.