MIT and Microsoft optimize agentic workflows to cut energy

Researchers from MIT and Microsoft have developed Murakkab, an intelligent system that automates the design and runtime optimization of agentic AI workflows to reduce computation, energy, and cost.

• Main announcement: Murakkab automatically maps developer intent to workflow components, selects existing models and tools, determines parallelism vs. sequential execution, and dynamically configures hardware and resource allocation at deployment to meet user constraints (e.g., prioritize accuracy or latency). The paper describing Murakkab will be presented at the USENIX Symposium on Operating Systems Design and Implementation and the research was supported in part by the Semiconductor Research Corporation and the U.S. Defense Advanced Research Projects Agency.
• Key results and details: In tests on video Q&A and code-generation workflows, Murakkab used about 35% of the computation compared with other methods, consumed about 27% of the energy, and ran for less than 25% of the cost; in one case it reduced energy consumption by more than an order of magnitude with only ~2% drop in accuracy. The system also provides cloud providers visibility across workloads so they can share computational resources more efficiently.