Build Agentic Workflows with OpenAI GPT OSS on SageMaker AI and Bedrock AgentCore

TL;DR

• OpenAI released two open-weight models, gpt-oss-120b (117B parameters) and gpt-oss-20b (21B parameters), both with a 128K context window and a sparse Mixture of Experts (MoE) design.
• The post demonstrates deploying gpt-oss-20b on SageMaker AI managed endpoints using a vLLM container, and orchestrating a multi-agent stock analyzer with LangGraph, then deploying to Amazon Bedrock AgentCore Runtime.
• A three-agent pipeline (Data Gathering Agent, Stock Performance Analyzer Agent, Stock Report Generation Agent) runs across Bedrock AgentCore while GPT-OSS handles language understanding and generation via SageMaker AI.
• 4-bit quantization (MXFP4) reduces model weights to 63 GB (120B) or 14 GB (20B), enabling operation on single H100 GPUs, with deployment options including BYOC (bring-your-own-container) paths and fully managed hosting through SageMaker AI.
• The solution emphasizes serverless, modular, and scalable agentic systems with persistent memory and workflow orchestration, plus clear steps for deployment, invocation, and cleanup. See the AWS blog for details. This article is based on the approach described by AWS and OpenAI in their documentation and demonstrations for building agentic workflows with GPT OSS on SageMaker AI and Bedrock AgentCore. AWS blog