OpenAI Open-Sources Codex Agent Harness Under Apache-2.0, Boosting ARC-AGI-3 Score from 13% to 38%
Summary
- • OpenAI open-sourced the Codex agent harness on August 19, 2026 under Apache-2.0 at github.com/openai/codex — the core execution system powering Codex App, CLI, and IDE extensions
- • Released components include the Codex CLI (codex exec), official Codex SDK, and app-server for persistent conversations, streamed events, and approval workflows
- • Harness optimizations — retained reasoning and context compaction — raised GPT-5.6 Sol's ARC-AGI-3 benchmark from 13.3% to 38.3% while reducing output tokens by 6×
- • The GitHub repository reportedly exceeded 100,000 stars shortly after launch, signaling strong developer adoption
Details
Codex agent harness open-sourced under Apache-2.0 on Aug 19, 2026
OpenAI published 'Codex as a platform: build on the open agent harness' on developers.openai.com on August 19, 2026, releasing the core execution system and agent loop that powers Codex App, CLI, and IDE extensions.
Components: Codex CLI, official SDK, and app-server released
Released components include: Codex CLI (codex exec for non-interactive tasks), the official Codex SDK, and the app-server handling persistent conversations, streamed events, and approval workflows — all at github.com/openai/codex.
Harness manages state, tools, sandboxing, and approval workflows
The agent harness handles conversation state, tool invocation, streaming execution, sandbox enforcement, approval workflows, and multi-turn context; developers can adapt it for custom products while the host application controls operational boundaries and tool access.
ARC-AGI-3: 13.3% → 38.3% — nearly 3× improvement via harness
Harness optimizations including retained reasoning and context compaction raised GPT-5.6 Sol's ARC-AGI-3 benchmark score from 13.3% to 38.3%, demonstrating that the orchestration layer itself is a major driver of agent benchmark performance.
Output tokens reduced 6× by the same harness optimizations
The retained reasoning and context compaction techniques that boosted ARC-AGI-3 performance also reduced output token usage by a factor of six, substantially cutting API costs for developers building on the harness.
100,000+ GitHub stars shortly after launch
The github.com/openai/codex repository reportedly exceeded 100,000 stars shortly after launch — an unusually strong early reception signal for developer infrastructure tooling.
Open harness, proprietary model — platform expansion strategy
OpenAI shares the infrastructure layer (harness, CLI, SDK) while keeping model weights proprietary — enabling an open developer ecosystem around Codex that increases API usage without giving away the core model assets.
Details compiled via Grok live web research from OpenAI and outlets
Core details sourced from OpenAI Developers Blog (August 19, 2026), Open Source For You (August 21, 2026), and GitHub via Grok live web research.
Sources: OpenAI Developers Blog (August 19, 2026); Open Source For You (August 21, 2026); GitHub; Grok Research.
What This Means
OpenAI's decision to open-source the Codex agent harness is a strategic platform move: sharing the infrastructure layer that orchestrates agent execution invites developers to build on Codex as a foundation while keeping model access through the API. The near-3× ARC-AGI-3 improvement (13.3% → 38.3%) with 6× fewer output tokens is the headline technical result, showing that the harness architecture itself — not just model capabilities — dramatically affects performance and cost. With 100,000+ GitHub stars and Apache-2.0 licensing, this has the hallmarks of a rapidly adopted developer standard, positioning OpenAI's open-source tooling as infrastructure for the broader agentic AI ecosystem. It also sets a competitive precedent: open-sourcing agent infrastructure to grow ecosystem lock-in is a playbook other frontier AI labs will likely consider.
Sentiment
Excited about harness improvements and open infrastructure, focused on practical gains over model scaling
“OpenAI open-sourced Harness yesterday - the execution engine behind Codex, released under Apache-2.0. With Harness, GPT-5.6 Sol's score on ARC-AGI-3 went from 13.3% to 38.3%, and token consumption dropped sixfold at the same time. Not a tradeoff - both numbers moved in the same direction. That's the part worth sitting with. Most of the AI cost conversation this year has been about which model to buy. This is a case where the harness around the model - how it plans, retries, and checks its own steps - did more for the bill than switching models would have.”
“Six times fewer output tokens. Same model. Same task. That's what a better harness does — and it's why OpenAI just open-sourced the engine behind Codex. The number that matters: on ARC-AGI-3, the exact same GPT-5.6 Sol model went from 13.3% to 38.3% just by changing the harness design. Retained reasoning and context compaction. While cutting output tokens sixfold. The model didn't improve. The system around it did. The agent loop is becoming open infrastructure.”
“OpenAI's Codex as a platform post gets the direction right: stop bolting chat onto everything, put the agent loop inside the product you already have. Harness tuning alone took Sol from 13.3% to 38.3% on ARC-AGI-3 with 6x fewer tokens. Worth doing the same for your own product.”
“OpenAI 把 Codex 的"发动机"开源了：不止 CLI，整个 Agent Harness 都放出来了... harness 设计本身对 Agent 能力的影响可能比换模型还大。已经有人在用了：Cisco 在 Cloud Control 的 App Builder 里用了 Codex SDK... 对开发者的实际意义：你不用再自己造 Agent 运行时了。”
Split
~90/10 positive on the value of open harnesses vs model-only scaling; little disagreement visible.