Using Codex and Claude Code Together on Amazon Bedrock: Implementation via Harness Engineering

In the first half of 2026, the terminal-based AI coding agent market has been reorganized around two axes: Anthropic's Claude Code and OpenAI's Codex. The two tools each possess different strengths—multi-agent orchestration versus high token efficiency.

Claude Code is based on the Claude Opus 4.8 model and boasts layered multi-agent workflows and excellent code defect detection capabilities. Codex, on the other hand, is primarily powered by GPT-5.5 (Spud) and offers Rust-based speed, OS-level sandboxing, and overwhelming token efficiency.

When combining these two agents, Amazon Bedrock plays a core role. Through Bedrock, authentication, governance, and cost tracking for both models can be integrated into a single AWS account, enabling a collaborative environment where you can freely switch models simply by modifying configuration files.

This experiment does not simply compare the intelligence of the models, but focuses on designing the Harness that connects the two agents. Through this, various metrics such as cost, time, and feature scores are measured to explore the optimal AI collaboration workflow.