DeepSeek has put the spotlight on the most unusual trick in its open-source agent’s book. On November 2 the company published a demo video and seven example prompts for Creator Mode, the Harness feature that lets users describe a plugin in plain language and watch the agent write the code, install it and check that it works. The mode itself is not new — it shipped as an experimental feature in the v0.2 preview in late September — but the showcase is the first time DeepSeek has shown it off in full. It comes as Harness, still a developer preview, draws some of the fastest adoption of any open-source agent project this year.
## What Creator Mode actually does
Creator Mode is a preset in Harness’s session menu, not a separate product or a new model. It arrived as an experimental Creation mode in the v0.2 preview in late September, and runs the same agent loop with extra powers: the agent can inspect the running app, write plugin code, install the result and verify that it works.
The demos DeepSeek published show the range. In one, a single request produces a floating Pomodoro timer in about five and a half minutes, built, installed and tested by the agent itself. Others include a design workspace, a poster tool and a paper-reading aid, each assembled from a plain-language description.
The remaining demos push further into interface building: a dashboard of conversation statistics, a sandbox game and a three-dimensional view of a transformer computing its next word. DeepSeek has published all seven demo prompts in Chinese and English, so users can rerun them, though it cautions that a single request will not always reproduce the video result.
## Built on “everything is a plugin”
Creator Mode is possible because of how Harness is built. DeepSeek describes the app as an agent harness in which tools, model providers, skills, memory and visual panels are all plugins rather than fixed code. That design means an agent that can write a plugin can change almost anything about the workspace it runs in.
The project has grown fast. Its public repository has passed 270,000 stars after opening in mid-August, and DeepSeek tuned its DeepSeek-V4.1-Flash-1017 model specifically for Harness, recommending version 0.2.1-rc.1 or later. The harness itself costs nothing to use; what users pay for is model usage through DeepSeek’s API, billed by the token.
## The safety warning matters as much as the demo
DeepSeek is blunt about the risk. Harness is a developer preview, its safety notice says the software has not undergone a security audit, and plugins are code: model-written code can run commands, damage files or [leak credentials](https://geroundtable.com/microsoft-artex-github-azure-secrets/) if a prompt or plugin goes wrong. Breaking changes are promised, so plugins built today may need rework as the platform matures.
## Our Read
The shift is not that an agent can write code — coding agents do that all day — but that the agent can change its own home. A plugin that adds a panel or a tool is a small thing; a loop in which each plugin makes the next job easier is the shape of a platform, not a chatbot.
DeepSeek’s framing, ‘Your Harness. You Define It,’ captures the bet: the workspace itself is now programmable by the person using it. The comparison that matters is with the closed rivals. OpenAI is [buying its way into agent coding workflows](https://geroundtable.com/openai-buys-pierre-git-agent-coders/), and [Claude Code](https://geroundtable.com/cat-wu-openai-claude-code-lead/) remains the polished incumbent; DeepSeek is answering with openness, letting anyone inspect, fork and extend the tool.
Openness cuts both ways, though. The same plugin system that lets a user add a timer lets a bad prompt add something worse, and DeepSeek’s own warning — no audit, breaking changes ahead — is the honest price of moving this fast.
The test now is whether the plugin ecosystem grows into something durable, with shared, reviewed plugins, or stays a collection of demos. DeepSeek says an official plugin marketplace is planned, which would be the next real signal to watch for the project.
## Why It Matters
If agents can extend themselves, the moat moves from the model to the ecosystem around it. DeepSeek is betting that an open, plugin-shaped agent, free to download and cheap to run on its own API, can outgrow closed rivals by letting users build the features they actually need.
For developers, the practical question is narrower: whether plugins built in Creator Mode survive the breaking changes DeepSeek has already warned about, and whether a shared plugin marketplace turns one-off demos into tools other people can trust and reuse safely.
## Background
DeepSeek released Harness in August as an open-source agent harness under the MIT licence, and shipped desktop apps with the version 0.2 preview in late September. Shortly before the showcase, it published a design article on ‘super-agent’ systems and open-sourced a prototype that runs several tasks in parallel, its own tests claiming half finished inside 12.4 minutes on synthetic workloads.
Source: [DeepSeek’s Harness page](https://www.deepseek.com/en/harness/) and [Quartz’s report](https://qz.com/deepseek-creator-mode-custom-ai-plugins-110226)
0 Comments