OpenAI Open-Sources the Codex Agent Harness
CodexOpenAI open-sourced Harness, the Rust-based execution engine behind Codex's CLI, SDK, and app-server, releasing it under the Apache-2.0 license on August 20, 2026. The move lets developers embed Codex's own agent loop — task memory, tool calls, sandboxing, and approval workflows — directly into their own products instead of building on top of a generic chat interface. OpenAI says harness-level optimizations alone lifted GPT-5.6 Sol's ARC-AGI-3 score from 13.3% to 38.3% while cutting token usage sixfold, and cited Cisco, Thrive Holdings, and Crete as early adopters already running production agents on it.
Featured Video
A video we selected to help illustrate this changelog
Key Takeaways
- OpenAI rebuilt Codex's harness around WebSocket connections instead of HTTP round-trips after GPT-5.3 Codex Spark's Cerebras-powered ~1,000 tokens/second made network latency, not model inference, the new bottleneck.
- Nearly every feature that makes Codex distinctive — tool search,
apply_patch, WebSocket streaming, and context compaction — is exposed through the Responses API, so any compatible harness can use them. - The harness is written in Rust and released under Apache-2.0, making the exact execution loop OpenAI runs internally available for developers to fork and modify.
- OpenAI's engineers argue that harness design, not raw model capability, is now the primary lever for agent performance — a claim the ARC-AGI-3 results are meant to back up.
- The optimized harness alone lifted GPT-5.6 Sol's ARC-AGI-3 score from 13.3% to 38.3% while cutting token usage sixfold, showing execution-loop engineering can rival model upgrades in impact.
- Enterprises including Cisco, Thrive Holdings, and Crete are already running production agents built on the harness, from a cloud App Builder to processing 7,000 tax returns.
Sources & Mentions
5 external resources covering this update
Codex, Behind the Harness — Dominik Kundel, OpenAI
YouTube
Harness engineering: Leveraging Codex in an agent-first world
Hacker News
Exclusive: OpenAI's secret weapon underneath Codex
The Deep View
Codex CLI: OpenAI's Open Agent Harness — Installation, Commands, and CI Integration
SitePoint
OpenAI Open Sources Codex Harness Framework
Open Source For You
What Shipped
On August 20, 2026, OpenAI open-sourced Harness, the execution engine that has quietly powered every version of Codex — the CLI, the ChatGPT desktop app, the IDE extension, and Codex's cloud service — under the permissive Apache-2.0 license. Three components make up the release: codex exec, a CLI entry point built for scripts, CI jobs, and one-off automated tasks that runs a bounded agent loop and returns structured output; the official Codex SDK, a programmatic interface for starting, resuming, and streaming agent tasks from an application; and app-server, the engine behind persistent conversations, streamed events, tool exposure, and human-in-the-loop approvals. Until this release, these pieces underpinned Codex's own products without being positioned as a general-purpose framework other developers were meant to build on top of.
Why the Harness Matters More Than the Model
OpenAI's pitch is that harness design — not just the underlying model — is what actually determines how well an agent performs. The company pointed to results on the ARC-AGI-3 benchmark, where retained reasoning and context compaction inside the harness lifted GPT-5.6 Sol's score from 13.3% to 38.3% while cutting output-token usage sixfold. That gain came entirely from how the execution loop manages memory, tool calls, and context over long-running tasks, not from a model upgrade.
Built for Embedding, Not Just Chat
By open-sourcing the harness, OpenAI is inviting developers to skip the generic chatbox interface and wire Codex's agent loop directly into their own products — engineering tools, operations dashboards, financial software, or anything else that needs task comprehension, tool invocation, sandboxed execution, and approval workflows built in. OpenAI cited early adopters already doing exactly that: Cisco built an App Builder into its cloud control platform using the Codex SDK, while Thrive Holdings and Crete embedded the harness into tax-preparation workflows, processing 7,000 returns and cutting preparation time by roughly a third.
Reception
The announcement drew a large and skeptical Hacker News discussion (297 points, over 200 comments), with commenters debating whether harness-level throughput gains translate into genuinely better software or just favorable-looking benchmarks. In a companion conference talk, OpenAI's Dominik Kundel traced part of the redesign to Cerebras-powered inference reaching roughly 1,000 tokens per second on GPT-5.3 Codex Spark — fast enough that HTTP round-trip latency, not model inference, became the harness's new bottleneck, prompting a rebuild around WebSocket connections. Kundel's talk also noted that most of what makes Codex distinctive — tool search, apply_patch, WebSocket streaming, and context compaction — is exposed through the Responses API, meaning any compatible harness can use them, not only OpenAI's own.