A keyboard for the agent fleet
OpenAI has released a 230 dollar keyboard for running Codex agents, with status lights, shortcut keys and a dial for how hard the model thinks. The gadget is a novelty. The instinct behind it, that a fleet of coding agents needs a control surface, is exactly right, and we already run one.
On 15 July OpenAI put a keyboard on sale. The Codex Micro costs 230 dollars, was co-designed with a specialty keyboard maker, and is pitched as a command centre for agentic work. It has light-up Agent Keys that show the status of each running agent, customisable Command Keys for frequent Codex actions, a joystick for launching workflows, and a dial that sets how much reasoning the model applies to a task. It is a limited run, so read it as a statement piece rather than a product, per TechCrunch.
The easy take is that a 230 dollar keyboard for talking to a chatbot is peak hype. That take is mostly correct and slightly lazy. The interesting part is not the object, it is the admission underneath it: that once you are running several coding agents at once, you need somewhere to watch them and something to press.
The moment you stop being a typist
For years the keyboard was where you did the work. You wrote the code. The machine waited. A control surface for agents flips that. The agents write the code and the human watches, redirects, and approves. The dial for reasoning effort and the lights for agent status are quietly honest about the new job: you are managing a set of processes, deciding which one needs more thinking time, and catching the one that has gone quiet or wandered off. That is a supervisor's console, not a writing desk.
We know the shape of that job because we do it. This studio runs its own fleet of coding agents daily, on real work, not demos. So the question the Codex Micro raises is one we have already had to answer in practice: what does a person actually need in front of them to run agents well.
What the control surface is really for
Three things, and none of them need a bespoke keyboard. First, status you can trust at a glance: which agent is working, which is blocked, which finished and is waiting on a decision. Second, a fast way to intervene: to interrupt an agent heading down a wrong path before it burns an hour, because the cost of a confident mistake compounds when nobody is watching. Third, and least glamorous, ownership: a clear record of what each agent changed, so a human signs off on what ships rather than discovering it later.
The dial that sets how hard an agent thinks is the whole idea in miniature. The work is no longer writing the code, it is deciding how much judgement each task deserves.
Hardware is the wrong layer
A physical console makes sense when the thing you are controlling is physical, a mixing desk, a lighting rig, a machine on a floor. Agents are not. They live in software, they scale past the number of keys you could sensibly light up, and the useful controls change every week as the tooling moves. Bake the control surface into a keyboard and you have frozen a fast-moving problem into plastic. The right place for it is the software you already run the agents from, where status, interruption and an audit trail can keep pace with the work.
How we read it
The Codex Micro is a novelty, and it will date quickly. But it is a useful tell. It says out loud that serious use of coding agents is becoming a supervision task, and supervision needs a surface: something that shows you state, lets you step in, and holds a record. If you are wiring agents into real work, that is the part to build carefully, and it is worth far more than the accessory. We would spend the 230 dollars on better observability, not a keyboard.