Moonshot AI launches Kimi Work, a desktop agent that runs a swarm of 300 subagents

Moonshot AI, the Chinese startup known for its Kimi chatbots, has released Kimi Work, a local desktop agent for macOS and Windows. Reports indicate it runs on the Kimi K2.6 model, and its architecture can run up to 300 subagents at once. It is one of the first commercial "agent swarm" products aimed at a broad audience rather than just developers.

Agent swarm architecture

Kimi Work installs directly on the user's computer and runs locally, without constantly sending data to external servers. The system is built around a multi-agent swarm: a lead orchestrator agent takes a task, breaks it into subtasks, and hands them off to hundreds of specialized subagents that work in parallel. Each subagent handles a specific type of operation: searching the web, analyzing text, interacting with web interfaces, or structuring and saving results.

Thanks to parallel processing, the system can potentially finish in minutes what a sequential agent would take hours to complete. The stated ceiling of 300 concurrent subagents is one of the highest figures for any consumer AI product to date. User data is processed on the device and is not sent to the cloud by default.

That sets Kimi Work apart from most agent services, which run through an external API.

WebBridge: controlling the real browser WebBridge is one of the product's core modules.

Many agents open an isolated browser and make you log in again on every site. WebBridge instead connects to the browser the user already has running, with existing sessions. Logins, cookies, and browsing history stay untouched, so the agent sees pages exactly as the user does.

That lowers the barrier to entry considerably: there's no need to set up separate integrations or hand passwords over to the app. The agent works wherever the user is already logged in, including work email, corporate tools, news services, and online stores. Kimi Work also includes a background task scheduler: the user queues a job, moves on to something else, and the agent keeps working in the background.

Key features: Controls a logged-in browser through WebBridge with no need to sign in again Runs up to 300 subagents in parallel for multi-step tasks Task scheduler for background execution without user involvement Local data processing with nothing sent to the cloud * Supports macOS and Windows out of the box ## The Kimi K2.6 model Kimi Work is reportedly built on Kimi K2.6, an updated version of Moonshot AI's flagship model.

The base K2 model posted competitive results on agentic benchmarks, especially in planning, coding, and multi-step reasoning tasks. K2.6 appears to be optimized for long agentic chains and stable parallel operation.

The company has not yet published official technical details on the new version, and some of the details come from third-party observations.

What it means

Until now, agent swarm products have mostly existed as API tools for developers or closed enterprise solutions. Kimi Work is betting on the mass market: install the app and you get hundreds of parallel agents that research, collect, analyze, and structure information without constant human involvement. The stated subagent count far exceeds anything competitors have announced publicly. How well that translates into real-world performance will only become clear through hands-on use.