Moonshot AI Launches Kimi Work — Desktop Agent with Swarm of 300 Sub-Agents

Moonshot AI, a Chinese startup known for its Kimi chatbots, has launched Kimi Work — a local desktop agent for macOS and Windows. According to available information, the Kimi K2.6 model runs under the hood, and the architecture allows simultaneous deployment of up to 300 sub-agents. This is one of the first commercial products in the "agent swarm" category aimed at a broad audience, not just developers.

Agent Swarm Architecture

Kimi Work is installed directly on the user's computer and operates locally — without constant data transmission to external servers. The system is built on the principle of a multi-agent swarm: a master orchestrator agent receives a task, decomposes it into subtasks, and distributes them to hundreds of specialized sub-agents that work in parallel. Each sub-agent is tailored to a specific type of operation: information search on the web, text analysis, interaction with web interfaces, structuring and recording results.

Due to parallel processing, the agent system can potentially handle in minutes tasks that a sequential agent would complete in hours. The declared limit of 300 simultaneous sub-agents is among the highest figures for a consumer AI product today. User data is processed on the device and by default is not transmitted to the cloud — this is a fundamental distinction of Kimi Work from most agent services that operate through an external API.

WebBridge: Managing the Real Browser

One of the product's key modules is WebBridge. Unlike agents that open an isolated browser and require re-authentication on each site, WebBridge connects to the user's already-running browser with preserved sessions. Logins, cookies, and browsing history remain intact — the agent sees pages exactly as the user sees them. This significantly lowers the barrier to entry: no need to set up separate integrations or hand over passwords to the application. The agent operates where the user is already authenticated: in work email, corporate tools, news services, and online shops. Kimi Work also supports a background task scheduler — the user queues a task, switches to another activity, and the agent continues working in the background.

Key capabilities:

• Managing an authenticated browser via WebBridge without re-login
• Parallel execution of up to 300 sub-agents for multi-step tasks
• Task scheduler — background execution without user involvement
• Local data processing without cloud transmission
• Support for macOS and Windows out of the box

Kimi K2.6 Model

Kimi Work is presumed to be powered by the Kimi K2.6 model — an updated version of Moonshot AI's flagship development. The base K2 model showed competitive results in agent benchmarks: particularly in planning, coding, and multi-step reasoning tasks. The K2.6 version appears to be optimized for long agent chains and stable parallel operation. The company has not yet published official technical details about the new version — some details are based on third-party observations.

What This Means

Until now, products with agent swarm architecture existed primarily as API tools for developers or closed corporate solutions. Kimi Work bets on the mass market: install the application and get hundreds of parallel agents that research, gather, analyze, and structure information without constant human involvement. The stated number of sub-agents significantly exceeds what competitors have publicly named. How much this corresponds to actual performance will become clear with practical use.