MIT, Microsoft optimize agent workflows with 65% computation cut

MIT and Microsoft Cut Agent Costs With Murakkab Optimization System

MIT and Microsoft researchers have developed Murakkab, a system that automatically optimizes how AI agents execute workflows by selecting the most efficient models, tools, and hardware configurations for each task MIT News.

The system's name—Urdu for "a composition of things"—reflects its core function: composing agentic workflows with minimal computational overhead. Published June 25, 2026, the research addresses a critical inefficiency in modern AI operations: agent workflows often run full-scale models and tool chains even when simpler, cheaper alternatives would satisfy user requirements.

How Murakkab Optimizes Agent Execution

Murakkab uses Mixed Integer Linear Programming (MILP) to simultaneously optimize four competing objectives: latency, cost, accuracy, and energy consumption. Rather than applying one large model to every step, the system intelligently routes tasks to appropriately-sized models and selects which tools to invoke—dramatically reducing waste.

In benchmarking, Murakkab met user requirements while using only about 35 percent of the computation required by traditional methods—a 65% reduction. Energy consumption dropped to 27 percent of baseline levels, a 73% reduction, for less than 25 percent of the cost MIT News. In one instance, energy dropped by more than 10 times with only a ~2% drop in accuracy, demonstrating that significant efficiency gains need not sacrifice quality.

Timing and Research Venue

The timing reflects growing industry pressure to reduce the operational costs of agentic AI systems. As enterprises scale agent deployments, per-task costs compound rapidly. Murakkab's approach—treating workflow optimization as a discrete mathematical problem rather than a sequential pipeline decision—offers a fresh architectural angle.

The research was presented at OSDI 2026 (Operating Systems Design and Implementation), the premier conference for systems research. The work emerged from collaboration between researchers at MIT's Cambridge campus and Microsoft's Redmond offices.

Implications for Agent Deployment

The system addresses a concrete pain point: current agentic frameworks often default to "use the best model for everything," wasting compute on tasks that don't require frontier capabilities. Murakkab's optimization layer sits between the agent's task scheduler and the execution runtime, making decisions about resource allocation transparent and quantifiable.

This is not a roadmap or concept—the system has been implemented, benchmarked against real agent workflows, and published with reproducible findings. For enterprises evaluating AI agent platforms, Murakkab represents a shift from "how capable is this agent?" to "how efficiently can this agent use compute to meet my requirements?"