Source: OpenAI Developers — 2026-08-20
Summary
OpenAI open-sourced Harness, the execution engine that has been powering its Codex coding agent, under an Apache-2.0 license — meaning developers can now build their own agent products on the same core instead of being limited to OpenAI's chat interface. The release bundles the codex exec CLI, an SDK, and the app-server that runs the actual agent loop: task comprehension, long-conversation memory retention, tool invocation, interruptibility, status syncing, and human-in-the-loop approval. OpenAI's headline claim is that optimizing this harness layer alone — with no change to the underlying model — lifted GPT-5.6 Sol's score on the ARC-AGI-3 benchmark from 13.3% to 38.3%, while cutting token usage sixfold.
Key Takeaways
- License and scope: Apache-2.0, includes
codex exec(CLI), the Codex SDK, and app-server — the full execution loop, not just a thin wrapper around API calls. - Headline benchmark: GPT-5.6 Sol went from 13.3% to 38.3% on ARC-AGI-3 purely from harness-level changes (retained reasoning and context compression), with 6x lower token consumption — the model weights never changed.
- The core loop it exposes handles the unglamorous but critical parts of running an agent in production: interruptibility mid-task, status synchronization across a long session, and human-in-the-loop approval gates for risky actions.
- Positions "harness quality" as a distinct, measurable lever separate from model quality — a framing increasingly common across the industry (see also DeepSeek's own harness release this month) as labs compete on the execution layer, not just raw model capability.
- Developers can now embed the same engine OpenAI uses for Codex into their own products, rather than building an agent loop from scratch on top of the raw chat API.
Reel Script
Hook: OpenAI just proved something that should worry anyone benchmarking AI purely on model size: the same model nearly tripled its score on a hard reasoning benchmark without a single weight being retrained.
Core Concept: When people talk about an AI "agent," they usually picture the model — GPT, Claude, whatever. But the model is only half the system. The other half is the harness: the code that decides what the model sees at each step, how much of the conversation history gets kept versus summarized, when to retry a failed tool call, and how to compress a sprawling multi-hour task down into something that fits in context. OpenAI just open-sourced that harness — the same one running Codex — so developers can build on it directly instead of reinventing this plumbing themselves.
Hands-On: The number that matters here is 13.3% to 38.3%. That's GPT-5.6 Sol's score on ARC-AGI-3, a benchmark designed to be hard for pattern-matching and to actually require reasoning — before and after OpenAI optimized the harness around it, with zero changes to the model itself. Two techniques did most of the work: retained reasoning, where the agent keeps its intermediate reasoning steps available across turns instead of discarding them, and context compression, where the harness aggressively summarizes what's no longer immediately relevant so the model isn't drowning in stale context. The side effect of doing this well: token usage dropped 6x, meaning the harness-optimized version was both dramatically smarter on this benchmark and dramatically cheaper to run.
Takeaway: If your team is agent-building and all your effort is going into prompt engineering or picking a bigger model, you're optimizing the wrong layer — the harness is now a first-class, benchmarkable piece of the stack, and OpenAI just handed you a reference implementation for free. Worth pulling apart before you build your own from scratch.