AI agent workflows are transforming how complex tasks are automated by chaining multiple models and tools into a single system. These workflows power applications like video analysis and real-time code generation, but their fragmented design often leads to over-provisioned resources, wasted energy, and inflated costs.
Now, a team from MIT and Microsoft has introduced an intelligent system that streamlines both the design and deployment of these workflows. Instead of requiring developers to manually specify every technical detail upfront, the new approach allows them to describe their goals in plain language. The system then automatically selects the optimal combination of AI models, tools, and hardware configurations to meet performance, cost, or energy priorities.
The challenge of static workflow design
Traditional agentic workflows demand extensive upfront configuration. Developers must choose which AI models to use, how to sequence tool interactions, and which hardware to allocate—all while navigating tradeoffs between speed, cost, and accuracy. This process is not only time-consuming but also prone to suboptimal decisions, especially as new models and tools emerge regularly.
“Even with careful planning, manually optimizing a workflow is nearly impossible,” says Gohar Chaudhry, lead author and an electrical engineering and computer science graduate student at MIT. “The configuration space is simply too vast, and black-box models from different providers add another layer of complexity.”
Cloud providers face an additional hurdle: they lack visibility into workflow internals, making it difficult to allocate hardware efficiently at runtime. This often results in over-provisioned resources that drive up both costs and energy consumption.
A self-optimizing system for real-world deployment
The new system, named Murakkab (meaning “a composition of things” in Urdu), shifts the burden from developers to automation. By accepting high-level descriptions of desired outcomes—such as “build a video Q&A app that extracts key frames, transcribes speech, and answers user questions”—Murakkab handles the rest.
It automatically composes the workflow by selecting the best-performing models and tools, determining execution order, and optimizing parallel processing where possible. The system also adapts dynamically to hardware and model advancements without requiring developer intervention.
Quantified efficiency gains in real-world tests
When evaluated on video Q&A and code generation tasks, Murakkab delivered consistent performance while significantly reducing resource usage. Across diverse workloads, it cut computational requirements by 65%, energy consumption by 73%, and operational costs by 75% compared to traditional methods.
“Agentic workflows are becoming the backbone of cloud services, and energy efficiency is no longer optional,” says Ricardo Bianchini, senior author and Microsoft Azure technical fellow. “Murakkab enables cloud providers to maximize efficiency while meeting user constraints like latency or cost targets.”
The team includes Adam Belay, an MIT EECS associate professor and MIT CSAIL member, along with additional researchers from Microsoft Azure. Their findings will be presented at the USENIX Symposium on Operating Systems Design and Implementation.
Looking ahead: smarter, greener AI at scale
As AI agents take on more complex roles in cloud environments, systems like Murakkab address a critical need for sustainability and cost control. By shifting optimization from static planning to dynamic adaptation, the approach not only reduces waste but also future-proofs workflows against rapid technological change.
For developers and cloud providers alike, the message is clear: efficiency doesn’t have to come at the expense of flexibility or performance. With the right automation, AI-powered systems can run smarter, cleaner, and more affordably than ever before.
AI summary
MIT ve Microsoft araştırmacıları, ajan tabanlı yapay zekâ iş akışlarını otomatik olarak optimize eden Murakkab sistemini tanıttı. Enerji tüketimini yüzde 73 azaltan bu yenilik, bulut bilişimin geleceğini şekillendirecek.
