Overview

OpenAI’s Gabriel Chua clarifies the confusing terminology around “Codex” by explaining it as a software engineering agent with three components: Model + Harness + Surfaces. The key revelation is that Codex models are specifically trained to work with their harness, making tool use and execution loops native behaviors rather than add-on features.

Key Points

  • Codex is defined as OpenAI’s software engineering agent, not just a model - it’s an integrated system that can execute tasks autonomously
  • The three-part structure (Model + Harness + Surfaces) means the harness contains instructions and tools, while surfaces are the interaction methods - this separates the AI capabilities from how users access them
  • The harness is open source in the openai/codex repository - developers can examine and potentially modify the tool collection and instructions
  • Codex models are trained specifically with their harness present - tool use and execution loops are native behaviors, not afterthoughts
  • This training approach means the harness evolves based on how the model operates - creating a co-designed system optimized for software engineering tasks