AI-powered workflows, or ‘agentic workflows’, are at the forefront of technological innovation. These systems have the power to chain together various models and external tools to tackle intricate tasks, such as analyzing a video and providing information about its content. However, they’re not without issues. Their design and deployment are often fragmented, leading to inefficiencies, waste, and unnecessary costs.
But all hope isn’t lost. Researchers from MIT and Microsoft have their heads in the game. They’ve developed an intelligent system that does double duty – streamlining the design of agentic workflows and optimizing their implementation on its own.
This new take on AI workflow is innovative, to say the least. Developers can describe their desired outcomes in plain language, without getting bogged down in the details. The system then takes the reins, autonomously deciding which models and tools to use. It also neatly handles all the technical details like hardware configuration and resource allocation when a cloud provider is running the show.
What’s more, this system isn’t static. It adjusts these configurations based on what the user wants at any given moment. Want to keep costs down? It’s on it. Need lightning speed? It adjusts accordingly. Testing shows the system dramatically reduces the computational requirements, making energy usage and costs far more efficient without compromising on performance.
Cooking up agentic workflows is no walk in the park. These systems consist of several autonomous AI agents using various models and tools to complete multi-step tasks like data processing and code generation. They’re the unsung heroes behind many user-facing applications we use every day.
Enter Murakkab. Taking its name from the Urdu word for ‘composition of things’, Murakkab optimizes the entire agentic workflow process. Developers describe what they want their application to do, and Murakkab identifies the best models and tools for the job.
Murakkab doesn’t just identify the best players, it makes the game plan. It decides which components should run sequentially and which can side-step delays by running in parallel. It even adjusts for changes in the field, flexibly responding to emerging models or graphic processing units (GPU accelerators).
When it comes time for cloud providers to deploy the application, Murakkab stays on top, making sure the workflow aligns with the user’s constraints like accuracy and latency. It figures out the best deployment schedules and hardware allocations, all while maintaining high levels of efficiency.
Even better, Murakkab doesn’t just work in theory, the testing shows it slashes computational use to 35% of what other methods require. It also consumes about 27% of the energy other methods do for less than 25% of the cost.
Looking to the future, researchers are already thinking about how to expand Murakkab for use with more complex workflows and larger computing clusters. Plans are underway to explore new optimization opportunities for agentic applications. This research, generously supported by the Semiconductor Research Corporation and the U.S. Defense Advanced Research Projects Agency, shows promise in making AI workflows much more resource-efficient, especially on major cloud platforms like Azure and AWS.
For a deeper dive, check out the original news article. And if you’re looking to introduce AI automation to your company, take a look at implementi.ai to see how they can revolutionize your business operations.
This website uses cookies.