O conteúdo está em inglês; a navegação e os filtros mantêm o idioma selecionado.

DeepSeek Harness: Can an SDK Become an AI Data Gateway?

DeepSeek Harness is presented as an agent workspace that can run through web, terminal, headless and SDK interfaces. It may broaden the use of agents in real software projects, but the available evidence does not show that it automatically creates a DeepSeek-controlled data flywheel.

Publicado em 08/09/2026, 03:16:56Última atualização 08/09/2026, 03:16:56Publicação editorial

The available material describes DeepSeek Harness as a developer-preview environment designed to let agents work inside codebases, terminals and longer-running tasks. Alongside an interactive interface, it offers terminal, headless, ACP, JSON-RPC and Python SDK entry points, as well as plugins, task orchestration and multi-agent workflows. These interfaces could make the system easier to embed in enterprise processes rather than limiting it to a standalone coding product.

The data-flywheel argument, however, should be treated as a hypothesis rather than an established fact. The cited documentation says sessions can be stored locally in JSONL or SQLite, telemetry is off by default, and any enabled telemetry is routed according to the deployer’s configuration. Local traces generated through other models are not automatically sent to DeepSeek, so Harness usage alone does not demonstrate a continuous source of training data for DeepSeek.

GLM-5.3 is reported to be available through ZCode and other products, while post-training, tool use and longer task environments are described as important routes to capability gains. Still, the evidence does not independently establish the reported model scale, benchmark improvements, data feedback loop or revenue impact. Usage, conversion and margins remain the more meaningful tests of the business thesis.

Evidências das fontes 6

A verificação de fontes não garante exatidão permanente. Considere a fonte original, a data e as correções posteriores.

DeepSeek正式开源Harness：它终于有了自己的Vibe Coding入口_新浪财经_新浪网finance.sina.com.cn · Complementar

以官方目前开放的形态来看，DeepSeek Harness 更接近一套运行在用户本地环境里的 Agent 工作台。它不只给出答案，还可以进入实际工作环境继续行动。比如，用户可以把一个代码项目所在的文件夹交给它，让它先阅读项目结构和文档，再寻找相关代码、修改文件、调用终端运行程序或测试；如果测试报错，它还可以根据错误信息继续定位问题、再次修改，而不是每一步都等用户复制粘贴代码和报错信息。这些能力主要通过 Web UI 提供，同时还有面向终端用户的TUI、适合脚本和 CI 的 Headless 模式，以及 ACP、JSON-RPC 和 Python SDK 等自动化入口。换句话说，同一套 Harness 既可以做成人直接操作的编程工作台，也可以被接进自动化流程：例如收到一个任务后自动检查代码、执行测试，完成后再返回结果。提前体验资料显示，这些形态共享同一套模型、会话和底层插件，只是通过不同组件组合成不同的产品形态。它还支持更复杂的长任务和多 Agent 协作。一个主 Agent 可以把工作拆给多个子 Agent，例如让一个负责搜索项目和资料，一个负责修改代码，另一个负责执行测试，再由主 Agent 汇总结果并决定下一步。官方框架同时提供计划、目标、待办事项和后台任务等机制，目的就是让 Agent 不必局限在“一问一答”，而可以持续完成一串彼此关联的操作。这让 DeepSeek 开始进入 Claude Code、Codex 等产品已经率先展开的竞争。基础模型公司不再满足于提供一个 API，而是进一步争夺模型和真实计算环境之间的执行层。 [...] 这也是为什么，同一个模型放进不同 Harness，实际表现可能出现明显差异。提前体验 DSH 的媒体曾将同一版本 V4 Flash 放入不同 Agent 框架完成相同任务，结果出现了肉眼可见的差异。当然，单次测试无法证明哪种 Harness 普遍更强，但至少说明，工具、提示词、上下文组织和执行策略已经成为 Agent 最终性能的一部分。过去谈 DeepSeek，竞争焦点几乎始终落在模型本身：参数规模、训练成本、Benchmark 和 API 价格。但到了 V4 Pro 与 DeepSeek Harness，这条边界正在发生变化。模型仍然决定智能的上限，但当模型开始真正进入代码库、终端和长期任务后，如何把这