DeepSeek AI Releases DeepSeek Harness in Developer Preview: An MIT-Licensed Agent Harness Where Everything is a Plugin DeepSeek AI 发布 DeepSeek Harness 开发者预览版：MIT 许可的全插件化 Agent 框架

TL;DR

• DeepSeek released DeepSeek Harness v0.1 as an MIT-licensed developer preview, framing the architecture as Agent = Model + Harness
• The Cordis plugin kernel makes every component swappable — models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and UI are all configurable plugins without source code changes
• Four runtime modes (Standard, Code, Minimal, Creator) load different default plugin sets for varying use cases from full coding agents to benchmarking environments
• An append-only session log captures every context injection, enabling resume, fork, search, and replay on a single event stream — a more granular observability claim than typical agent frameworks
• Provider-agnostic by design, supporting DeepSeek, Anthropic, OpenAI, Bedrock, Vertex, Azure, Codex, and any OpenAI-compatible endpoint

Why It Matters

DeepSeek Harness represents a shift from fixed agent frameworks to composable, plugin-driven agent runtimes, which could become the default infrastructure layer for enterprise AI deployments. Its MIT license and self-hosted capability make it particularly attractive for regulated industries like finance and healthcare that require auditable, locally-deployed agent systems. The emphasis on full observability through append-only session logs addresses a critical gap in current agent frameworks where debugging and replay capabilities are limited.

Technical Details

• Cordis Kernel: A meta-framework based on "A Programming Paradigm for Spatiotemporal Composability" that handles plugin mounting, unmounting, and dependency management. Capabilities live entirely in plugins rather than a privileged core, enabling runtime composition without code changes.
• Plugin Architecture: Models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and UI all sit behind Cordis plugin boundaries. Any component can be selected, swapped, or extended through configuration alone.
• Four Runtime Modes: Standard mode provides full coding agent capabilities (file editing, shell, search, skills, planning, subagents). Code mode exposes tools via a TypeScript SDK for multi-step operations. Minimal mode strips to two tools (bash, str_replace_editor) for benchmarking. Creator mode adds runtime inspection and preset-authoring.
• Observability System: Every context injection, system prompt, reasoning trace, tool call, and subagent scheduling event is written to an append-only session log. The Trajectory view enables source-based inspection, while resume, fork, search, and replay operate on the same event stream.
• Model Routing: Provider-agnostic design with API key management stored in $DSH_HOME/.credentials.yaml. Supports DeepSeek, Anthropic, OpenAI natively, plus Bedrock, Vertex, Azure, and Codex with native credentials. Custom providers accept any OpenAI-compatible base URL.

Industry Insight

• Enterprise AI teams should evaluate Harness as infrastructure for building internal coding agents over private repositories, particularly in regulated sectors where auditability and local deployment are mandatory.
• The plugin-first architecture positions Harness as a potential reference implementation for the next generation of agent frameworks, suggesting that modularity and observability will become key differentiators rather than novelty features.
• The emphasis on benchmarking (Minimal mode) and run replay indicates DeepSeek is targeting both production agent development and model evaluation use cases, potentially establishing new standards for agent performance measurement.

TL;DR

• DeepSeek发布DeepSeek Harness v0.1开发者预览版，采用MIT许可证开源，定位为Agent运行时组装工具而非固定编码助手
• 基于Cordis元框架实现"一切皆插件"架构，模型、工具、技能、会话、沙箱、存储、循环、调度和UI均可通过配置动态替换
• 提供四种运行时模式（Standard/Code/Minimal/Creator），分别面向完整编码Agent、TypeScript多步操作、模型基准测试和插件开发场景
• 采用append-only会话日志记录所有上下文注入，支持轨迹回放、分支、搜索和恢复，实现完整的Agent可观测性
• 模型路由插件化设计，原生支持DeepSeek、Anthropic、OpenAI、Bedrock、Vertex、Azure、Codex及OpenAI兼容端点

为什么值得看

DeepSeek Harness重新定义了Agent框架的设计范式，将"模型+环境"的耦合层抽象为可插拔架构，为AI原生应用开发提供了灵活的基础设施层。其开源MIT许可和全链路可追溯特性，使其成为受监管行业（金融、医疗）和学术研究进行Agent部署与评估的理想选择。

技术解析

Cordis插件化架构：Harness运行在Cordis元框架之上，内核仅负责插件挂载、卸载和依赖管理，所有能力（模型、工具、技能、会话、沙箱、存储、循环、调度、UI）均以插件形式存在，开发者可通过配置文件选择、替换或扩展任意组件，无需修改源码。

四种运行时模式：Standard模式提供完整编码Agent功能（文件编辑、Shell、搜索、技能、规划、子Agent和工作流）；Code模式通过TypeScript SDK暴露工具组合能力；Minimal模式仅保留bash和str_replace_editor两个工具用于模型基准测试；Creator模式增加运行时检查和内存插件实验功能。

全链路可追溯性：所有模型可见内容（系统提示、推理过程、工具调用及结果、子Agent调度、上下文注入）均写入append-only会话日志，Trajectory视图支持按来源检查记录，恢复、分支、搜索和回放共享同一事件流。

模型路由插件化：通过Settings → Models界面配置，DeepSeek API密钥无需重启即可生效；内置目录支持Anthropic和OpenAI；AWS Bedrock、Vertex、Azure和Codex需要原生凭据；自定义提供商接受任何OpenAI兼容的base URL和协议。凭据以只写方式存储在$DSH_HOME/.credentials.yaml中。

部署与SDK：通过npx @deepseek-ai/dsh web启动Web UI（默认端口3600）；提供Python SDK（deepseek-harness-sdk，需Python 3.10+，支持Linux x64/arm64和macOS 14+ arm64），内置运行时无需系统Node.js。

行业启示

Agent基础设施层标准化趋势加速：DeepSeek Harness的"一切皆插件"设计反映了Agent框架从硬编码向可组合架构演进的必然方向，未来企业级Agent部署将更依赖此类可插拔基础设施而非封闭产品。

可观测性成为企业级Agent刚需：append-only日志和完整上下文追踪的设计，回应了金融、医疗等受监管行业对Agent决策可审计性的核心需求，预示可追溯性将成为Agent框架的标配能力。

模型无关性降低供应商锁定风险：原生支持多提供商路由且凭据分离存储的设计，使企业能够灵活切换底层模型而不重构Agent运行时，为AI原生应用开发提供了更强的架构韧性。