This page may contain stale information. Last updated: 2026-06-28

Definition

Murakkab (Urdu: “composition”) is a resource-efficient serving system for agentic-ai workflows developed by MIT CSAIL and microsoft Azure Research. It decouples workflow specification from execution configuration via a declarative abstraction.

Key Points

  • Profile-guided optimizer + adaptive runtime for model/hardware selection
  • USENIX OSDI 2026 paper (June 2025 publication date)
  • Benchmarks: up to 2.8× GPU, 3.7× energy, 4.3× cost reduction while meeting SLOs
  • Video Q&A and code generation: ~35% compute, ~27% energy, <25% cost vs baselines

Sources