DeepSeek Harness autonomous coding agents

This dossier tracks DeepSeek Harness coding-agent prototypes, their permissions and usage limitations.

Last updated : 2 November 20262 key moments

DeepSeek logo, identifying the developer; copyright © 2023 DeepSeek, MIT license. — Copyright © 2023 DeepSeek / Wikimedia Commons — MIT https://opensource.org/license/mit ; permission d’utiliser, copier, modifier et distribuer sous réserve de conserver la notice de copyright et la permission ; fourni sans garantie. Rendu PNG.

Where things stand

As of 2 November 2026, the Harness team has demonstrated Creator Mode and published seven prompts for creating plugins through conversation. DeepSeek says some results required multiple exchanges; the earlier autonomous prototype remains research code.

2 November 2026Reported information

DeepSeek demonstrates conversational plugin creation mode

The DeepSeek Harness team demonstrated Creator Mode on 2 November 2026, enabling natural-language requests to create plugins. It published seven demo prompts in Chinese and English for users to try recreating the examples.

Examples cover a research-paper reader, poster workbench, video editor, 3D visualizations and conversation-statistics dashboard. The official page instructs users to copy a prompt into Harness and select Creator Mode.

DeepSeek says some plugins were refined over multiple exchanges, so a single instruction may not exactly reproduce the video result.

These are demonstrations and shared prompts, rather than evidence that every generated tool is reliable or ready for professional use. The consulted sources provide no independent evaluation of these results.

Sources 2

31 October 2026Reported information

DeepSeek releases prototype agent for GitHub development tasks

DeepSeek opened the code for dsh-superagent-prototype on 31 October 2026, a coding agent built around DeepSeek Harness. Released under MIT, it watches one GitHub repository and can also receive requests through Feishu.

DeepSeek logo, used to identify the company. — Copyright © 2023 DeepSeek / Wikimedia Commons — MIT https://opensource.org/license/mit ; permission d’utiliser, copier, modifier et distribuer sous réserve de conserver la notice de copyright et la permission ; fourni sans garantie. Rendu PNG.

When an authorized person requests work, it opens a work item, executes it in its own session and directory, then replies in the original conversation, for example with a pull request. A dashboard highlights matters needing human attention.

The team describes it as a research prototype, adapted for publication and different from its internal version. Some described ideas are not implemented; production use therefore requires review and adaptation.

Sources 3