Sarms paper claims GraphWorkflow runs compiled agent graphs up to 62.5x faster than LangGraph

KyeGomezB · x · 2026-09-08

The Swarms team published a systems paper on GraphWorkflow, the graph execution engine inside the open-source Swarms framework, built on a compile-once, execute-many model: an optimized execution plan is built up front, avoiding the repeated cost of interpreting and constructing large agent graphs.

Key results vs LangGraph 1.0.4, across 15 topology/size configurations (10–200 nodes), medians of 9 samples with 95% confidence intervals:

The paper includes a formal execution model with proofs, a cost taxonomy for agent orchestration, engine design details, a programming-model comparison with LangGraph, and full benchmarks. The DAG programming model supports parallel execution, branching, fan-in/fan-out, loops, and dependencies. The benchmark harness is open, and every number is claimed to be reproducible.

Related event: Swargs Open-Sources GraphWorkflow, Up to 62.5x Faster Than LangGraph(10 posts)→

Original post →

More from coding & agent

coding & agent channel →