After Kimi's large model frequently breaking performance rankings, Moonshot AI has made another move - officially open-sourcing the new AI agent development framework Kosong, providing a lightweight, flexible, and highly scalable underlying support for building next-generation intelligent agent applications. The framework is centered around the core concept of "simplifying complexity and unleashing creativity." Upon its release, it sparked heated discussions in the developer community, seen as a crucial bridge connecting large models with real-world application scenarios.

image.png

Unified Abstraction Layer, Say Goodbye to "Patchwork" Development

Current AI agent development often faces challenges such as tool fragmentation, incompatible interfaces, and difficult debugging processes. Kosong addresses these issues by building a unified LLM abstraction layer, encapsulating different large models, tool functions, and memory modules into standardized components. Developers no longer need to rewrite logic for each model. Its core innovation lies in an asynchronous tool orchestration engine, which efficiently schedules multiple tools to run in parallel. For example, it can simultaneously call a weather API, map service, and calendar system to complete complex task flows like "Is it going to rain at 10 AM tomorrow? If so, please change it to an indoor café," significantly improving the agent's response efficiency and task completion rate.

Plugin-Based Design, Completely Free from Vendor Lock-In

Kosong explicitly supports plug-and-play chat providers (Chat Provider), allowing developers to freely switch between different models such as Kimi, DeepSeek, and Llama3, or even use them together without modifying the core business logic. This design greatly enhances the flexibility of technology selection, preventing teams from being tied to a single AI service provider and also reserving a seamless upgrade path for future model iterations.

Out-of-the-Box, Python3.13+ uv for Rapid Onboarding

The framework is built on modern Python ecosystems, requiring Python 3.13 or higher and recommending the use of the next-generation package manager uv for installation, which is several times faster than pip. The official provides complete examples: from calling the Kimi model and implementing streaming responses to creating custom asynchronous tools (such as addition functions and database queries), with clear and concise code. The built-in demonstration agent allows users to run it with just setting environment variables, experiencing a complete AI agent workflow within 5 minutes.

AIbase believes that the true value of Kosong lies in transforming AI agent development from a "high-barrier engineering" process into a modern software practice that is "composable, reusable, and iterative." When developers are no longer burdened by underlying compatibility issues and can focus on business logic and user experience innovation, the explosive growth of intelligent agent applications may accelerate. As an important part of Moonshot AI's "open source + open" strategy, Kosong is not just a set of tools, but an invitation - an invitation to global developers to jointly build a smarter and freer AI future.

Project address: https://github.com/MoonshotAI/kosong