OpenAI open-sourced codex harness on August 20 under Apache-2.0. The release turns the core framework behind Codex into code developers can modify and commercialize. It also pushes AI agent loops into software built for real work instead of a generic chat window.
OpenAI and Codex Harness
Harness is not a model. It handles task understanding, long-conversation memory, relevant-information review, tool use, progress display, exception handling, human approval requests at critical points, and structured results.
That design explains why OpenAI says the framework can live inside products, engineering tools, operations dashboards, and financial software. The openai/codex repository on GitHub includes three core components, and the app-server is the piece that keeps conversation state, streams events, interrupts work mid-stream, exposes proprietary tools, and routes approval requests.
ARC-AGI-3 and ARC Prize
OpenAI also published performance data tied to Harness design. On ARC-AGI-3, retaining reasoning and context compression lifted GPT-5.6 Sol from 13.3% to 38.3%, and the same changes cut output token count to one-sixth of the original.
That result sits uneasily beside the ARC Prize official standardized score of 7.8% for GPT-5.6 Sol. François Chollet said that general features available to all API users can fall within a reasonable scope when they are not benchmark-specific optimizations.
Thrive Holdings and Crete
OpenAI has already pointed to outside use. Thrive Holdings and Crete embedded Harness into professional tax preparation workflows, processed 7,000 returns, and cut tax preparation time by roughly one-third while incorporating real-time feedback from tax professionals.
Cisco used the Codex SDK to build an App Builder within its cloud control platform. The open question is which three core components sit in openai/codex and what each one does, because that is the part developers need before deciding where to adopt Harness first.
