Can I handle errors in specific parallel branches without failing the whole graph?

Yes, the skill includes patterns for error isolation using asyncio.gather with return_exceptions=True to ensure one failing branch doesn't terminate the entire workflow.

How do I aggregate results from multiple parallel nodes?

The skill demonstrates using Python TypedDicts with Annotated types and the operator.add reducer to automatically accumulate list-based results from parallel branches.

How does the Send API work in this skill?

The Send API allows the graph to dynamically trigger multiple instances of a specific node (fan-out) based on the current state, which then automatically converge (fan-in) upon completion.

What is the benefit of parallel execution in LangGraph?

Parallel execution significantly reduces total latency by running independent tasks or multiple agent calls concurrently rather than sequentially.

Does this skill support timeouts for parallel tasks?

Yes, it provides implementation patterns for setting per-branch timeouts to ensure that a single slow worker doesn't block the completion of the entire graph.

LangGraph Parallel Execution

Name: LangGraph Parallel Execution
Author: yonatangross

byyonatangross

•

データサイエンスとML

Implements high-performance parallel execution patterns in LangGraph to enable concurrent agent workflows, fan-out/fan-in structures, and efficient map-reduce operations.

This skill provides production-ready implementation patterns for running independent nodes concurrently within LangGraph. It equips developers with standardized approaches for the Send API, fan-out/fan-in architectures, and map-reduce workflows, significantly reducing latency in multi-agent systems. By incorporating advanced concepts like error isolation, branch-specific timeouts, and state accumulators using Annotated reducers, it ensures that complex parallel workflows remain robust and manageable even when individual branches encounter issues.

主な機能

01Send API implementation for dynamic, parallel branch generation

02Standardized Fan-Out/Fan-In patterns for task distribution and aggregation

03Map-reduce patterns for processing large datasets concurrently

04Advanced error isolation and per-branch timeout management

0569 GitHub stars

06State accumulation using Annotated reducers for seamless data merging

ユースケース

01Building robust data processing pipelines where failures in one branch don't stop the entire graph

02Scaling a single task across multiple LLM calls to reduce total execution time

03Running multiple specialized agents simultaneously to analyze a single input

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add yonatangross/orchestkit langgraph-parallel

For use in Claude.ai and ChatGPT

Download Skill