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 license2. The release gives developers the ability 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 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 the impact of the Harness orchestration layer on model performance. With Harness optimization, GPT-5.6 Sol's score on the ARC-AGI-3 benchmark rose from 13.3% to 38.3%, while token consumption dropped sixfold. ANALYSIS The benchmark delta — nearly tripling the score while simultaneously reducing token consumption — underscores that agent-level orchestration design, not just model scale, is a significant performance lever.
Two enterprise deployments are already live. Tax partners used the framework to process 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 is framed as a competitive move. The open-sourcing marks an escalation in OpenAI's competition with Anthropic in the computer-use agent space, according to the announcement coverage. OpenAI's team is addressing safety governance challenges through confirmation policies that balance security and usability.
ANALYSIS By releasing Harness under Apache-2.0, OpenAI is making its agent orchestration stack a potential default for third-party developers — a distribution play that shifts competition from proprietary agent tooling toward model-level differentiation. The permissive license allows commercial use without restriction, lowering the barrier for startups and enterprises to build agent applications on OpenAI's architecture rather than constructing their own from scratch.
The ARC-AGI-3 results are notable for what they attribute to the harness rather than the model itself: the same GPT-5.6 Sol checkpoint produced dramatically different scores depending on the orchestration layer, reinforcing the emerging view that agent scaffolding is a first-class engineering surface alongside model training.