Gohar Chaudhry was staring at a tangled web of AI models and tools—each operating in isolation, consuming excess energy, and slowing down critical applications—when he and his team at MIT and Microsoft had a breakthrough. They envisioned a smarter way to build and run agentic workflows, the complex AI systems that power everything from video analysis to code generation. The result is Murakkab, a system named after the Urdu word for "a composition of things," that’s redefining how these workflows are designed, deployed, and optimized.
Agentic workflows are increasingly central to cloud computing, stitching together multiple AI agents, models, and external tools to complete multi-step tasks. But their fragmented nature has long been a bottleneck. Developers typically hard-code every component and configuration in advance, leaving little room for adaptation and often leading to over-allocated resources, higher costs, and unnecessary energy use. With AI demand surging, these inefficiencies are no longer sustainable.
Murakkab changes the game. Instead of requiring developers to specify every technical detail, it lets them describe the application’s goal in plain language—like building a video Q&A tool that extracts key frames, generates transcripts, and answers user questions. From there, Murakkab automatically selects the best models and tools, determines which tasks can run in parallel, and configures the optimal hardware setup. When deployed in the cloud, it dynamically adjusts resource allocation based on user priorities, whether that’s speed, cost, or energy efficiency.
In real-world tests, Murakkab slashed the number of computational units needed to run agentic workflows, cutting energy use and costs significantly without sacrificing performance. For cloud providers like Microsoft Azure, where efficiency directly impacts both sustainability and profitability, this is transformative. "Enabling a cloud provider to intelligently make these workflows more resource-optimal is a win for everyone involved," says Chaudhry, an EECS graduate student at MIT and lead author of the paper. He co-authored the work with Adam Belay, associate professor of EECS and MIT CSAIL member, and senior author Ricardo Bianchini, technical fellow and corporate vice president at Microsoft Azure.
Presented at the USENIX Symposium on Operating Systems Design and Implementation, Murakkab represents a shift toward adaptive, intelligent system design. As new models and hardware emerge, the platform evolves with them—no manual reconfiguration needed. For developers, this means faster innovation. For users, it means faster, greener AI. And for the planet, it’s a step toward making the AI revolution not just smarter, but more sustainable.