OpenAI announced on August 20 the full open-sourcing of Harness, the core execution framework powering its Codex coding agent, released under the Apache-2.0 license4,3. The permissive license allows developers to freely modify and commercialize the framework, embedding AI agent loops into their own products, engineering tools, operations dashboards, and financial software.
Harness is the execution system that drives Codex's AI agents, handling task comprehension, memory retention, tool invocation, and human approval workflows. The open-source release comprises three components: the codex exec CLI tool, the Codex SDK, and the app-server1. The framework supports persistent conversations, real-time streaming, and human approval workflows.
OpenAI published benchmark data showing that optimizing Harness design alone boosted GPT-5.6 Sol's score on the ARC-AGI-3 benchmark from 13.3% to 38.3%, while cutting token consumption sixfold. ANALYSIS The benchmark gain — nearly tripling the score — paired with a sixfold reduction in token usage suggests the framework's orchestration layer contributes meaningfully to agent performance independent of the underlying model weights.
Two enterprise deployments are already in production. Thrive Holdings integrated Harness into a tax preparation workflow that processed 7,000 returns, reducing preparation time by roughly one-third. Cisco built an App Builder on its cloud platform using the Codex SDK.
The release positions Harness as a direct competitor to Anthropic in the computer-use agent space, according to the reporting. OpenAI's team is addressing safety governance challenges through confirmation policies that balance security and usability.
ANALYSIS By open-sourcing the orchestration layer rather than the model itself, OpenAI retains control over its frontier weights while commoditizing the agent execution stack — a structure that could drive adoption of OpenAI's API as the default model backend for Harness-based applications. The Apache-2.0 licensing removes friction for commercial integrators, contrasting with more restrictive approaches to model-weight distribution.