Kimi Agent
The task mode behind OK Computer, swarm, Work, and Claw.
A direct definition
Kimi Agent is the mode where Kimi treats your request as a job with steps, not as a question that ends in one paragraph. Kimi's help center says the first version of that mode shipped on 26 September 2025 under the name OK Computer. The agent is described as building its own approach, picking tools, and adjusting while the task runs.
A search for Kimi Agent in 2026 can mean four doors that share that idea. The web and the app run it in chat. Agent Swarm splits one job across many subagents. Kimi Work runs it on your computer, next to your files. Kimi Claw runs an always-on agent in the cloud. The help page lists all of these as part of the agent story, not as rival brands.
What a run is allowed to do
On the web, the agent help describes research, long documents, and tool use around material you upload or that is already online. Kimi Work extends that to local files, commands, and the browser extension. The extension help page says an agent can open pages, click, fill forms, and pull information, using the browser you already signed into.
That is also the boundary. An agent that can click your logged-in browser can change real accounts. Read the plan before you approve a step that sends mail, pays, or deletes. Kimi Desk cannot see or stop a run inside Kimi. The product's own confirmations are the control.
One task, written so an agent can run it
A usable agent request names the output and the stop condition. 'Read these three PDFs and return a one-page brief with a bullet for each open question' can finish. 'Look into the market' cannot. Attach the files or point Kimi Work at the folder. Say what the agent must not do: do not email anyone, do not buy anything, do not delete files. The help center's agent story is about tool choice. Your job is to bound the tools.
If the work is a pile of similar items (fifty companies, a folder of contracts), that is the case Agent Swarm is documented for: split the pile, then merge. If it is one paragraph of Python, you do not need a swarm. Use a normal chat or the code prompt shape. Swarm spends more of the credit pool because more runs happen.
Credits and the model underneath
Agent features share a membership credit pool, metered by tokens. Kimi's membership overview lists website deployment, Deep Research, Slides, Kimi Code, Kimi Work, and Kimi Claw on that pool, and it says Kimi Code also has its own 5-hour weekly limit that does not apply to the other features. Unused credits expire at the end of the billing cycle. The dollar amounts and credit counts are on pricing.
The model names (K2, K2.6, K2.7 Code) are easy to mix up with the agent feature. The model is the weights. The agent is the loop that calls tools. Kimi K2 separates those names. You can call the model from your own software through the API without opening the Kimi Agent button at all.
Questions
- Is Kimi Agent the same as OK Computer?
- OK Computer is the name Kimi's help center uses for the agent mode launched on 26 September 2025. Current help pages fold that behavior into Kimi Agent, Agent Swarm, Kimi Work, and Kimi Claw.
- Does every account get Agent Swarm?
- On the K2.6 pricing table Kimi published, Agent Swarm is marked on the paid tiers and not on the free Adagio row. Concurrent subagents increase on higher tiers. Check the live table before you buy.
- Can Kimi Agent control my browser?
- Browser control is the Kimi browser extension (formerly WebBridge), used with Kimi Work or another local agent. Ordinary web chat does not click through your existing logins by itself.
- Where is the official overview?
- Kimi's help center article 'What Is Kimi Agent?' is the primary source for the timeline and the entry points. This page quotes that article and links it below.
Keep reading
Sources
Checked 2026-09-25. Prices and limits can change on the live help pages.