OpenAI Open-Sourced the Codex Harness: Same Model, New "Base," Nearly 3x the Score
📖 Glossary
AI Box (also known as Agent Computer / Agent PC), is a dedicated local hardware device that runs AI Agents. Pre-installed with an AI agent management system, plug-and-play, running 24/7. Users can remotely command AI to work via Discord, Slack, Telegram, WhatsApp, and more.
Abstract: On August 19, OpenAI open-sourced the execution engine behind Codex — the Codex Harness — under Apache-2.0 at github.com/openai/codex. The core idea is simple: Agent = Model + Harness. With the same GPT-5.6 Sol model, changing only the Harness (not the model) lifted the ARC-AGI-3 score from 13.3% to 38.3% while cutting token usage sixfold. This article explains what open-sourcing the Harness actually changes — and what it doesn't: whether your Agent has a 7×24 online "home" is something open source can't solve.
On August 19, OpenAI made a move on its official blog that rattled the developer community more than a new model release would have: it open-sourced the underlying execution engine that powers Codex — the Codex Harness — under the Apache-2.0 license, with the repository live at github.com/openai/codex.
The reaction was louder than any model launch. Not because the model got smarter — quite the opposite. It made people suddenly realize that what really decides whether an AI Agent "can work and works well" is often not the model itself.
1. What Exactly Is the Harness: the Agent's "Base"
Many people still understand an AI Agent as "a good model + a good prompt." Anyone who has run long-horizon tasks knows that's far from the truth.
A usable Agent is not just a prompt and one model response. It needs to understand the task, retain memory across a long conversation, inspect relevant information, call tools, show progress, handle failures, pause to request human approval when necessary, and finally return a useful result. This whole execution system around the model is the Harness.
OpenAI itself gave a very direct formula: Model + Harness = Agent. The model is responsible for "thinking"; the Harness is responsible for turning thinking into doing.
The heart of this release is the Harness's three-layer integration interface:
- codex exec: non-interactive command-line invocation, ideal for unattended scenarios like CI/CD pipelines, returning structured output directly;
- Codex SDK: TypeScript and Python programmatic interfaces, letting developers control task start, pause, resume, and streaming via code;
- app-server: embeds the Agent into your own product through a JSON-RPC protocol, supporting persistent sessions, real-time event streams, and approval callbacks.
In other words, OpenAI handed over the entire low-level capability of "how to make an Agent run, run stably, and run for a long time."
2. The Most Striking Number: Same Model, Nearly 3x the Score
The most telling figures in the announcement are ones OpenAI published itself: on the ARC-AGI-3 benchmark, the very same GPT-5.6 Sol model, with only two Harness-level optimizations — retained reasoning and context compaction — jumped its score from 13.3% to 38.3%, while total output tokens dropped to one sixth of the original.
No model upgrade. No parameter stacking. Just tuning the "base" lifted the score nearly threefold while slashing cost to one sixth.
What does this mean? It means many Agents fail or struggle not because the model isn't smart enough, but because of how it gathers context, calls tools, survives multi-hour tasks, and stays safe within boundaries. All of that lives in the Harness layer. The model sets the ceiling; the Harness decides whether your Agent ever reaches it.
So OpenAI's open-sourcing has a clear strategic intent: I may not give away the model for free, but I'll open the "base" to you. Build your product and workflows on my Harness, and naturally your model calls and budget stay in my ecosystem.
For ordinary people, this is actually an amplifier of good news. In the past, if you wanted an AI that automates work, you either waited for giants to feed you features or had the technical chops to tinker yourself. Now that the Harness is open source, the barrier drops significantly — you don't need to be a compiler expert to plug an Agent into your own business flow. Open source is democratization, and that deserves genuine applause.
3. But What's Open-Sourced Is the "Base," Not the "Home"
Amid the excitement, there's a point most people overlook yet is critically important: open-sourcing the Harness solves "how to do it," not "where to do it or when to do it."
No matter how beautifully you build your Agent, it has to run somewhere. Most people's first instinct is — run it on their own computer. The problem is obvious: when you clock off and shut down, the Agent "clocks off" too; when you travel, sleep, or go out for the weekend, it lies motionless inside the powered-off machine. At 3 a.m. when data needs monitoring or a report needs generating, your Agent is "asleep" inside that shut-down computer.
More realistically, even during the day it's hard to keep it running. Your computer has to work, attend meetings, play games, and run all kinds of software; the Agent can only squeeze into a corner consuming memory, and you constantly worry it'll clash with other programs. It's hard to free up a machine that is "dedicated, never off, always on standby."
This is the part an open-source framework truly cannot solve for you: a runtime environment that is always online and on standby. The Harness gives you a smart brain and flexible hands and feet, but it cannot give you a heart that keeps beating forever.
4. A Real Personal AI Assistant Is a Standalone Hardware Box
Once you see this clearly, a lot of things fall into place.
A true "personal AI assistant" was never a piece of software installed on your computer. Software, by its nature, can only boot and shut down with your computer. What truly deserves the name "assistant" should be a standalone hardware box:
- Plug it in and it's online 7×24; when you turn off your computer, lose connection, or travel, it keeps running;
- While you work during the day, it monitors data, collects emails, and organizes materials at home;
- While you sleep at night, it runs tasks off-peak, generates daily reports, and watches for anomalies;
- When you wake up the next morning, the deliverables are already there, viewable on WeChat.
Cloud AI can also be "always online," but it carries two inescapable costs: first, your data sits on someone else's servers; second, every task run is billed by usage. A local hardware box flips exactly those two points — data stays in your own home, bought once and run long-term. That is the premise on which "private" and "assistant" can both hold true at the same time.
It is also worth noting what the open-source Harness does not change about cost: the model itself is still accessed through OpenAI's API or a ChatGPT subscription. Open-sourcing the runtime does not make the model free. A local box pairs the free, adaptable Harness with a model of your choice, and you decide when to run expensive calls — for example, off-peak — keeping the bill under your control.
5. "Affordable to Build" Is Not "Able to Run"
Agent frameworks are increasingly open source and free — a trend across the whole industry and a dividend for ordinary people. But we must stay clear-headed: "affordable to build" and "able to run" are two different things.
Open-sourcing the framework solves "everyone can build"; always-on hardware solves "it truly keeps running." The former is the threshold, the latter is the environment. Lower the threshold, but if the environment doesn't keep up, the Agent you built is still a "half-finished product" — smart, but dead the moment power is cut.
So what really separates people is no longer "do you know how to use the framework." Anyone can learn a framework, and switching models is trivial. What truly separates them is: does your Agent have a 7×24 online home? Those with a home have an all-weather workforce; those without have only a small tool during computer-on hours.
6. Kaihe AIBOX Is That "Home"
At this point, the answer is self-evident.
Kaihe AIBOX is exactly such a standalone device: it stays online forever when plugged in, ships with OpenClaw and Hermes pre-installed, and works out of the box. You plug the open-source Codex Harness in, attach your model API, and it runs for you in the background 7×24.
The most critical point — data is stored in the local box, not uploaded to the cloud. What your Agent does and what it remembers is known only to you. That is the true meaning of "private."
Plug a brain like the Codex Harness into a box that never powers off, and your Agent truly "comes alive": working for you by day, standing night watch while you sleep. You sleep; it works.
This is the one thing most worth understanding in the open-source era: frameworks can be free, but the environment that lets AI truly work for you 24 hours a day is worth finding a home for.
Bring one home. Search "Kaihe AIBOX" on WeChat.