MIT, Microsoft Build Murakkab, Cutting AI Workflow Compute to 35%
Updated
Updated · MIT News · Jun 25
MIT, Microsoft Build Murakkab, Cutting AI Workflow Compute to 35%
1 articles · Updated · MIT News · Jun 25
Summary
Murakkab let MIT and Microsoft researchers turn plain-language app goals into optimized AI agentic workflows, then reconfigure models, tools and hardware in real time to match user priorities such as speed, cost or accuracy.
Tests on video Q&A and code-generation workloads showed the system met requirements using about 35% of the computation of other methods, about 27% as much energy and less than 25% of the cost.
The platform also decides which workflow steps should run sequentially or in parallel and gives cloud providers visibility across workloads, helping them share compute resources more efficiently.
In one case, Murakkab cut energy use by more than an order of magnitude with only about a 2% drop in accuracy, highlighting the tradeoff control the researchers say developers cannot manage manually.
The team plans to extend the system to more complex workflows and larger clusters as agentic systems become a bigger part of cloud platforms and their energy footprint.
Could making AI workflows 75% cheaper paradoxically increase the cloud's total energy consumption?
How will AI optimizers handle complex tasks where the 'best' outcome is subjective and not easily measured?
Does hiding complexity from developers create a new class of invisible, hard-to-debug AI system failures?
Murakkab: MIT and Microsoft’s AI Workflow System Achieves 3.4x Speedup and 4.5x Energy Efficiency
Overview
Murakkab, unveiled on June 25, 2026, is a new system developed by MIT and Microsoft Azure to tackle the inefficiencies and fragmentation in complex AI agentic workflows. By streamlining intricate AI operations, Murakkab achieves significant reductions in computational requirements, energy consumption, and operational costs, while maintaining high performance. Its standout feature is the ability to dynamically adjust configurations in real-time based on user priorities, making AI workflows more efficient and responsive. This innovation marks a major step forward in managing advanced AI systems, promising both sustainability and adaptability for future AI applications.