Powered by the Apache-2.0 licence, OpenAI’s open-source Harness engine slashes AI token consumption sixfold while boosting model benchmark scores.
On 20 August 2026, OpenAI officially open-sourced Harness, the core underlying engine and execution framework that powers its flagship coding agent, Codex. The framework was released under the Apache-2.0 licence, allowing developers to freely modify, embed, and commercialise it without being restricted to a general-purpose chat interface.
The release includes codex exec (CLI tool), the official Codex SDK, and app-server (the core engine execution server). Harness manages the execution loop for AI agents, including task comprehension, long-conversation memory retention, real-time event streaming, tool invocation, interruptibility, status synchronisation, and human-in-the-loop approval workflows.
Optimisation of the Harness framework alone (via retained reasoning and context compression) improved the GPT-5.6 Sol model’s score on the ARC-AGI-3 benchmark from 13.3% to 38.3%. Optimising the harness design reduced token consumption by sixfold (1/6th of the original output token volume), significantly lowering API call costs.
A tax-prep pilot using the framework processed 7,000 returns and reduced overall preparation time by approximately one-third. Enterprise organisations, including Cisco (for cloud management tools) and Thrive Holdings, have already deployed custom AI agents built on the Codex Harness.
