Source: MIT News — Researchers from MIT and Microsoft have developed Murakkab, an intelligent system that streamlines the design of AI agentic workflows and automatically optimises how those workflows are deployed in the cloud. The work will be presented at the USENIX Symposium on Operating Systems Design and Implementation.
Agentic workflows chain together multiple AI models and external tools — databases, Python programs, APIs — to complete complex multi-step tasks like analysing video or processing code. But these fragmented systems are notoriously inefficient, often over-allocating computational resources and wasting energy. Murakkab lets a developer describe what they want the workflow to do in plain language, then automatically selects the optimal models, tools, hardware configuration, and resource allocation.
"Agentic workflows are getting very complicated and quickly becoming the backbone of what cloud providers are doing. Energy usage is a huge concern," said lead author Gohar Chaudhry, an MIT EECS graduate student. In testing, Murakkab reduced the computational units needed for deployment without hampering performance — a win for both cloud costs and carbon footprints.
Analysis
Murakkab solves a problem most of the industry is still pretending doesn't exist. The current norm is to throw an LLM at every subtask in an agent pipeline — language model for planning, language model for retrieval, language model for tool selection — and hope the cloud bill works out. That's comically wasteful, and Murakkab's plain-language interface points to where the real value is: letting developers specify intent and having the system figure out the cheapest path to execution. Microsoft Azure co-authorship is the signal here — if Murakkab (or something like it) ships as an Azure optimisation layer, it could become the default for enterprise agent deployments. The side-effect few are discussing: a system that automatically picks the *cheapest* model for each subtask may systematically route around the frontier labs' most expensive offerings, which is great for efficiency and terrible for their per-token revenue models.
