Agentic workflows are artificial intelligence (AI) procedures that link multiple models and tools to complete complicated, multi-step tasks. Because these systems are broken into many parts, they often waste computing power, energy, and money. To solve this, researchers from MIT and Microsoft created a system called Murakkab. This system simplifies the creation of these workflows and improves how they run.
Usually, developers must hard-code technical choices and manually choose every model, tool, and piece of hardware in advance. This is difficult because the systems use black-box models, which are programs where the inner workings are hidden, and each has complex settings. Also, a cloud provider that offers computing services over the internet, cannot see inside the workflow to share resources efficiently. Murakkab fixes this by letting developers describe what they want the application to do in plain language. The system then selects the best existing models and tools, deciding which steps run in order and which run at the same time.
How the system improves computing
When the application runs for a user, Murakkab dynamically adjusts the setup based on what the user cares about, such as lowering costs or increasing speed. It figures out the best hardware configuration, including the use of a GPU accelerator for AI calculations. It adapts these choices in real time to maximize efficiency. The system also gives the cloud provider the ability to see multiple workloads and share computer power wisely.
In tests, Murakkab met user needs while using only about 35 percent of the computing power of older methods. It used about 27 percent as much energy and cost less than 25 percent of the price. The system can also make tradeoffs, like lowering energy use significantly in exchange for a tiny drop in accuracy. The researchers plan to expand this system to handle larger computer networks and more complex tasks in the future.
See also the preprint "Murakkab: Resource-Efficient Agentic Workflow Orchestration in Cloud Platforms."