DeepSeek and Peking University reveal the Harness self-evolution blueprint
- DeepSeek and Peking University published a paper titled "A Programming Paradigm for Spatiotemporal Composability," which the article says explains the core design behind Harness and its Cordis system.
- The paper centers on Cordis as a plug-and-play base that treats components as plugins that can be dynamically loaded, unloaded, and recombined.
- To address runtime changes without restarting the whole system, the paper introduces "revertible effects" for temporal composability and "reactive coeffects" for spatial composability.
- The article says the design has been used in the Koishi chatbot framework for four years and has been verified by more than 4,000 community plugins in production.
- The stated long-term goal is "self-evolution," where an agent can generate tools, install them into its runtime, and later replace them if needed.
DeepSeek and Peking University published a paper that the article describes as the technical blueprint behind Harness and its Cordis core. The paper, titled "A Programming Paradigm for Spatiotemporal Composability," focuses on how a runtime can support plugins that are freely plugged in, removed, and recombined.
According to the article, the main challenge is that ordinary plugin systems often require a full restart after unloading a component, which breaks accumulated context and state. The paper frames this as two problems: temporal composability, which concerns preserving system state across load and unload cycles, and spatial composability, which concerns managing dependencies between components without circular failures.
The proposed approach adapts ideas from type theory into the runtime layer. The article says the team introduces "revertible effects," where every state change records an inverse operation so unloading can restore the prior state, and "reactive coeffects," where components declare required dependencies and activate or deactivate automatically as those dependencies appear or disappear.
The article adds that this design has been running on the Koishi chatbot framework for four years and has been validated by more than 4,000 community plugins. Its stated goal is self-evolution: an agent that can generate tools for a task, install them into its own runtime, and later replace them if needed.
DeepSeek与北大披露Harness自进化蓝图
DeepSeek与北京大学发布了一篇论文，文章将其描述为Harness及其Cordis核心的技术蓝图。论文题为《A Programming Paradigm for Spatiotemporal Composability》，重点讨论如何让运行时支持插件自由加载、卸载和重组。
文章称，真正的难点在于，普通插件系统在卸载组件后往往需要整个平台重启，这会打断已有上下文和状态。论文把这一问题概括为空间可组合性和时间可组合性两部分，分别对应组件依赖关系的管理，以及插件加载卸载过程中系统状态的连续性。
文章介绍，方案将类型理论中的一些概念改造到运行时层面：通过“可回滚效应”，每次状态修改都记录逆操作，卸载时按相反顺序恢复；通过“响应式共效应”，组件声明自己需要的依赖，依赖满足时自动激活，依赖消失时自动停用并卸载。
文章还称，这套设计已在Koishi聊天机器人框架中运行四年，并经过4000多个社区插件的生产环境验证。其指向的长期目标是“自进化”，即代理能够按任务生成工具、安装进自身运行时，并在发现问题后自行替换。