OpenAI Engineer Clarifies Codex Agent Architecture And Training
On 22 February 2026, Gabriel Chua, OpenAI's Developer Experience Engineer for APAC, explains that Codex comprises a model, an open-source 'harness', and interaction 'surfaces'. He states Codex models are trained with the harness present, embedding tool use, execution loops, and verification into model behavior. The clarification signals that agent capabilities arise from co-training with tool-oriented runtimes, affecting how developers implement and extend Codex-based agents.
Key Points
- 1Defines Codex as Model + Harness + Surfaces, with Model+Harness forming the agent runtime
- 2States models are trained with the harness present, integrating tool use, execution loops, verification
- 3Notes harness is open-source in openai/codex repo, enabling developers to reproduce and extend agents
Scoring Rationale
Official insider confirmation increases significance; however, limited technical detail reduces immediate reproducibility and deep technical applicability.
Sources
Public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
