Murakkab: Resource-Efficient Agentic Workflow Orchestration in Cloud Platforms
Talkat 20th USENIX Symposium on Operating Systems Design and Implementation (OSDI '26) · 2026-07-13..07-15·
summary
Murakkab exposes agentic workflows as a declarative workflow spec and uses an LLM to pick agents then a profile-guided optimizer plus adaptive runtime to pick model, frame counts, hardware (A100/H100), parallelism and autoscaling. The optimizer formulates a mixed integer linear program over accuracy, latency, energy and cost using reusable agent, workflow and hardware profiles. Evaluation versus LangGraph shows up to 2.8× fewer GPUs, 3.7× lower energy and 4.3× lower cost while meeting query SLOs.
AI-written summary · about
chapters (9)
- 0:00Murakkab introduction
- 1:43Static workflow limits
- 3:26Cross-stack optimization
- 6:14Unified orchestration vision
- 8:41Profile-guided runtime
- 11:50Optimizer design MILP
- 13:33Dynamic cross-stack adaptation
- 13:54Experimental efficiency gains
- 15:59Summary and future
related talks
51mOptimizing Complex Workflows with Event-Driven Multi-Agentic Approach
50mUnpacking an Agentic Workflow
24mAccelerating Workflow Orchestration: Automated Code Gen Using Arazzo & MCP
50mHarnessing Event-Driven and Multi-Agent Architectures for Complex Workflows in Gen...
47mAgentic AI Patterns
31mAgentic AI architecture patterns
59mOptimizing Complex Workflows with Event - Driven - Multi Agentic Approach
22mAgentic AI for Engineering Leaders | Strategy & Execution Guide