OpenAI has open sourced Codex Harness, the execution framework that runs its Codex agents, and published the code under the Apache-2.0 licence. The company released the repository on GitHub and provided three ready components: the codex exec CLI, the official Codex SDK for TypeScript and Python, and the app-server execution engine. Greg Brockman reposted the announcement on X and described the move as widening Codex beyond programming tools.
OpenAI says Harness manages the agent loop, including task understanding, long conversation memory, tool invocation, real-time event streaming, interruptibility, human approvals and status synchronization. The company published data showing two Harness changes, retained reasoning and context compression, lifted the GPT-5.6 Sol model score on the ARC-AGI-3 benchmark from 13.3 percent to 38.3 percent. OpenAI also reported that the same optimisations cut output token volume by sixfold, reducing API call token consumption and associated costs.
Uses Reported And Implications For Developers
Early adopters have embedded Harness into vertical products and internal tools. OpenAI cited a tax-preparation pilot by Thrive Holdings and Crete that processed 7,000 returns and shortened preparation time by roughly one third. Cisco used the Codex SDK to add an App Builder into its cloud control platform. OpenAI demonstrated a Relay logistics dashboard that removes a chat box, feeds interface context to the agent, calls application MCP tools, requires human approval for risky actions and synchronises status back to the dashboard.
OpenAI and commentators argue the release hands interface, context and safety control back to application developers. The Apache-2.0 licence allows commercial modification and embedding, which OpenAI framed as a way to avoid forcing workflows into a generic chat interface. Developer reactions included a post by @ClusterProtocol praising the separation of front-end approval interfaces from the underlying execution loop. The company positioned Harness as a tool to embed agents into familiar dashboards, while letting hosts enforce data access and human gatekeeping.
