AgentCore Runloop integrates Codex for production agent workflows
In a Juejin AI article, author 寒斜 demonstrates an AgentCore application built with a Codex-based custom harness, covering model connections, MCP tools, skills, storage, identity, observability, and file-based multimodal handling. AgentCore Runloop connects Codex to platform services, while AgentCore Chat serves as the interactive front end. The author reports typical time to first token of about 1.7 seconds and a P95 of about 1.95 seconds across 20 samples, with no obvious stability issues or severe tail latency observed.
The article provides a detailed implementation and deployment reference for combining a Codex harness with AgentCore infrastructure, including security boundaries, tool registration, observability configuration, and persistence patterns.
作者：寒斜
本文章演示 AgentCore [ 1] 的自定义代码/镜像能力，目的是展示如何基于自定义的 Harness 框架（Codex） 结合 AgentCore 提供的身份、模型、工具、技能、凭证 、观测评估等能力实现一个端到端的 Agent 应用，内容包含两个部分，技术方案和实践教程，文末附相关源码和本次的测试数据，可以根据需要自行查阅。
技术方案
1、整体架构
Some source content is omitted. Continue reading on the original site.
#AgentCore#AgentCore Runloop#Codex#AgentCore Chat#Alibaba Cloud AgentIdentity#AgentLoop#OpenTelemetry#Qwen-VL-Max#Qwen-Image-Plus#Alibaba Cloud OSS#AgenticFS#MCP