Kimi OK Computer

Kimi OK Computer is the agent mode of Kimi, the AI assistant made by Moonshot AI. Rather than replying with text alone, it takes a request and works through the steps needed to produce a finished piece of work: a website, a slide deck, an organised data table. Kimi's help center dates its release to 26 September 2025 and calls it the first agent product mode in Kimi^[2]; the same arrival is described more generally as Kimi's agent mode launching in September 2025^[1].

Background: from a research agent to an agent mode

Kimi had already been given a task-specific kind of autonomy. In June 2025 Moonshot added Kimi-Researcher to the Kimi website and app, an autonomous research agent that carries a research task through on its own^[1]. OK Computer followed that and widened the scope from researching a question to producing the files in which a piece of work is delivered.

The model side matters too. Kimi's help center says Moonshot trained Kimi K2 with end-to-end reinforcement learning to be a native agent^[2]. "Native" here means the ability to plan a task and act on it step by step is trained into the model itself, rather than being supplied only by instructions wrapped around a chat model. For a user the difference shows up mid-task: whether the work keeps moving on its own or has to be carried forward step by step.

What it does

The help center lists website generation and PPT creation among the product lines of Agent mode^[2]. In use, the agent builds multi-page websites and editable slide decks from a prompt^[1]. Editable is the important word: the user receives a file they can keep working in, not a finished picture of a slide. Output is not limited to text — the agent also produces audio, images and video^[1].

An agent mode generally runs as a loop: the request is broken into steps, tools are called, the intermediate result is checked, and the work continues or is redone. That is what separates it from a single answer to a single question.

It also works on data. Wikipedia's entry says the agent can process up to 1 million rows of input data at once^[1]. Kimi's own help page describes a smaller case, listing support for organising 1,000 rows in Excel^[2]. The two figures are not the same kind of claim — a stated input capacity on one side, a documented example on the other — so the practical scale depends on what the task actually asks for.

Later: Agent Swarm

Later Kimi models added Agent Swarm, which runs hundreds of sub-agents in parallel^[1]. The pattern is division of labour: a large job is split into parts that several agents work on at the same time and whose results are then combined. In such systems the limit shifts from what one model call can produce to how well the work is divided and reassembled.

Where it sits

A chat assistant returns an answer to read; an agent mode returns a file to use. That changes what the user does. Instead of operating each tool in turn — laying out slides, wiring pages together, sorting a spreadsheet — they describe the outcome they want and then check the result. OK Computer is Kimi's general-purpose version of this: one mode covering several kinds of deliverable, rather than one specialised tool per task. It is also where the line that started with Kimi-Researcher, an agent built for a single job, extends into an agent mode meant for many.