Improving the speed and energy-efficiency of AI agents

· Source: MIT News - Artificial intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cloud Computing & IT Infrastructure, Robotics & Autonomous Systems · Depth: Intermediate, medium

Summary

A new system named Murakkab, developed by researchers from MIT and Microsoft, optimizes the design and deployment of complex AI agentic workflows. Introduced on June 25, 2026, Murakkab addresses inefficiencies in these multi-model, multi-tool systems, which often lead to wasted computation, energy, and cost. Developers can describe their application's intent in plain language, and Murakkab automatically identifies the best AI models, external tools, hardware configurations, and computational resource allocations. The system dynamically adjusts these settings based on user priorities like minimizing costs or maximizing speed. When tested on diverse agentic workloads, Murakkab reduced computational units needed for deployment, cutting energy requirements to approximately 27% and costs to less than 25% compared to traditional methods, all while maintaining performance. It achieved this by using only about 35% of the computation required by other methods.

Key takeaway

For AI Architects or MLOps Engineers designing complex agentic workflows, Murakkab changes how you approach system configuration. You should consider adopting systems that dynamically optimize model, tool, and hardware selections based on high-level intent. This can drastically cut your operational costs and energy consumption by up to 75% and 73% respectively, while maintaining performance. Evaluate platforms offering real-time resource allocation and adaptive configuration to streamline deployment and improve efficiency.

Key insights

Murakkab dynamically optimizes AI agentic workflows, significantly reducing computation, energy, and cost without performance loss.

Principles

Method

Murakkab takes high-level intent, identifies optimal models/tools, determines parallel/sequential execution, and configures hardware/resources dynamically for cloud deployment, meeting user constraints.

In practice

Topics

Best for: Machine Learning Engineer, NLP Engineer, Computer Vision Engineer, AI Engineer, MLOps Engineer, AI Architect

Related on AIssential

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by MIT News - Artificial intelligence.