Researchers from MIT and Microsoft Azure have built Murakkab, a system scheduled for presentation at OSDI 2026 that lets developers describe agentic AI workflows in plain language while automatically optimizing model selection, tool invocation, hardware scheduling, and energy use at deployment time. In reported tests, Murakkab used roughly 35% of the compute, 27% of the energy, and less than 25% of the cost of conventional hand-configured approaches, according to WindowsForum’s detailed analysis.
The Problem: Manual Wiring at Scale
Most agentic systems today are built the way early web applications were: developers choose a planner model, decide which tools it can call, stitch together retrieval and validation steps, specify execution order, and pick hardware tiers that seem likely to work. If latency is poor, they tune. If the bill is too high, they tune again.
That approach breaks down when workflows involve black-box models from multiple vendors, external tools with variable latency profiles, and customer-facing service-level objectives. A video question-answering workflow, for example, might need to extract frames, transcribe speech, summarize visual content, select relevant segments, and answer a user’s question. There is no single right implementation, just many configurations with different speed, accuracy, cost, and energy trade-offs.
How Murakkab Works
Murakkab separates what a developer wants from how the platform executes it. Instead of requiring hard-coded technical choices, the system accepts a higher-level description of the desired workflow. From there, it selects models and tools, infers dependencies, decides what can run in parallel, and generates an execution plan optimized against user-defined service-level objectives.
The system uses three components, as described in the WindowsForum report: a declarative abstraction layer for workflow specification, profile-guided optimization that learns how candidate configurations behave, and an adaptive runtime that adjusts execution dynamically. The name, drawn from an Urdu word meaning “a composition of things,” reflects the system’s view that agentic workflows are less like single applications and more like small, moving supply chains.
Microsoft’s Stake
Ricardo Bianchini, a Technical Fellow and Corporate Vice President at Microsoft Azure, is listed as senior author on the paper. That positioning signals this is not purely academic work. Microsoft has spent heavily to position Azure as an AI infrastructure platform, and the company’s broader AI strategy spans Copilot, Azure AI infrastructure, Microsoft Foundry, Fabric, GitHub, Windows, and enterprise security tooling. All of those become more complicated if agentic workloads are expensive, unpredictable, and difficult to schedule efficiently.
The cloud business incentive is straightforward: if agentic AI makes workloads more valuable but also more wasteful, the provider that can reduce waste without lowering user-visible quality gets an edge. In a market where GPUs remain scarce enough to shape product roadmaps, better orchestration is capacity strategy, not a green garnish.
From Bespoke to Compiled
Murakkab’s core argument is that bespoke agent deployment is inefficient at scale and the industry needs standardized abstractions for agentic optimization. The parallel to compilers in traditional software is deliberate: developers write high-level code, and the compiler handles optimization for the target hardware. Murakkab proposes the same relationship for agent workflows, where the developer specifies intent and the system handles execution efficiency.
The OSDI 2026 presentation will test whether that abstraction holds up under scrutiny from systems researchers.