← Back to the summary

DeepSeek Harness Is an Agent Framework Where Everything—Even the Agent Itself—Is a Swappable Plugin

The AI world was as lively as a festival last week!

First, DeepSeek dropped several bombshells in a row.

On August 13th, the official version of DeepSeek V4 Pro was released.

On the 14th, DeepSeek's own Agent product was also released: DeepSeek Harness.

Immediately after, Zhipu officially announced the release of GLM-5.3.

Then on the 17th, DeepSeek API prices rose sharply, from several times to over ten times.

Forget benchmarking, the update speed alone is too much to keep up with...

Today, let's focus on what everyone has been talking about recently: DeepSeek Harness.

Some might say: Damn, it's been out for a week, the hype is over, and you're only talking about it now?

Here's a unified reply: I never write articles just to chase trends. I only write about two kinds of things—what I find truly valuable, and what I see as a trend. Hype fades, but value doesn't. An Agent runtime framework like Harness will only become more important over time.

Furthermore, since last Friday, I've been busy rebuilding my own website and really haven't had the time these past few days, which is why this was delayed. 🥲

Alright, enough chatter. Writing isn't easy, so please hit that bookmark first, and let's get down to business.

First, About DeepSeek V4 Pro

Before talking about DeepSeek Harness, let's briefly touch on DeepSeek V4 Pro and the price hike.

This official version of DeepSeek V4 Pro mainly enhanced its Agent capabilities. The biggest shortcoming of the previously released preview version was Agent tasks, tool-chaining execution, and software engineering tasks. The official version focuses on filling this gap.

It also supports the Responses API and Codex access, and has been fully rolled out across the App, Web, and API platforms. (Of course, this isn't the focus of this upgrade, as these features were supported earlier.)

The Responses API is a next-generation interface standard launched by OpenAI for Agent intelligence. Now, DeepSeek V4 natively supports the Responses API, meaning it can directly connect to Codex without any protocol conversion, making the integration process much simpler than before.

What Does the DeepSeek Price Hike Mean?

On August 17th, DeepSeek officially announced a significant price hike.

Before the price hike:

After the price hike:

The cache hit price for V4 Pro rose from 0.025 yuan to 0.30 yuan during peak hours, a direct 12-fold increase.

The output price rose from 6 yuan to 27 yuan during peak hours, a 4.5-fold increase.

At this price, combined with the performance of DeepSeek V4 Pro, I can only say: the former "price butcher" has suddenly lost a lot of its appeal.

You see, DeepSeek's previous moat was, frankly, just one thing—cheap. For the same task, others charged full price while it offered a discount; where others discounted, it practically gave it away. Countless individual developers and small teams used it as their "cheapest main model," all for the sake of maximizing cost-effectiveness.

But after this price hike, things have changed.

In terms of performance, V4 Pro has indeed charged into the first tier, but domestic models like GLM-5.3 and Qwen 3.8 Max are no slouches either. These players were already squeezed onto the same level, with none able to decisively pull ahead of the others. Before, DeepSeek was "similar performance, but at a fraction of the price," making it a no-brainer choice. Now, with its price advantage shrinking, that "no-brainer" reason no longer holds.

Even from a comprehensive cost-effectiveness standpoint, looking solely at API calls, GLM might have already overtaken it.

So, what does DeepSeek's comprehensive price hike on August 17th signify?

I think this event is not just a price adjustment by one company; it's an inflection point signal for the domestic large model industry.

First, the supply and demand of computing power can no longer sustain the low-price model. Second, the logic of industry competition has changed, shifting from a race to the bottom on price to a competition on value. The playbook for the domestic large model market over the past year was: whoever is cheapest grabs the developers. Now, the logic has become: model capability first, price second.

A Focus on DeepSeek Harness

What exactly is DeepSeek Harness? How should we understand it? And how do we use it?

Next, let's explore these questions that newcomers care about most.

Many beginners might wonder why current AI Coding agents are all named something like XXX Code, such as Claude Code, OpenCode, Kimi Code. So why isn't DeepSeek's own product called DeepSeek Code, but DeepSeek Harness?

This is because DeepSeek Harness was never positioned to be built specifically for AI Coding.

Just as LangChain previously proposed when discussing Agent Harness, there is a very important formula:

Agent = Model + Harness.

A good Agent is necessarily composed of a Harness and a model together. (This formula is also mentioned on the DeepSeek official website.)

Many beginners might still not understand what a Harness actually is. Simply put, the Claude Code, Codex, Workbuddy, ZCode, etc., that we use are essentially all Harnesses. They must be paired with models like Claude Fable 5, GPT 5.6-Sol, GLM-5.3 to work; otherwise, they are just empty shells.

The internal structure of a Harness is actually quite complex.

Tools, Skill systems, session management, sandbox isolation, storage, Agent loops, sub-Agents, scheduling, workflows... a whole set of components, none of which can be missing.

The relationship between them, drawn as a diagram, looks roughly like this:

flowchart TB
    U["👤 User Input"] --> SM["🗂️ Session Management"]
    SM --> LOOP

    subgraph LOOP["🔁 Agent Loop (Core Engine)"]
        direction LR
        T1["🧠 Think<br>Model decides next step"] --> T2["⚡ Act<br>Invoke capability to execute"] --> T3["👁️ Observe<br>Collect execution results"] --> T1
    end

    LOOP -->|"Needs to work"| TOOLS["🛠️ Tools<br>File / Shell / Search"]
    LOOP -->|"Needs experience"| SKILLS["🎯 Skill System"]
    LOOP -->|"Task too big, split it out"| SUB["🤖 Sub-Agent"]

    TOOLS --> SANDBOX["📦 Sandbox Isolation<br>Execution environment safety zone"]
    SUB --> SCHED["📅 Scheduling"]
    SCHED --> WF["🔗 Workflow"]
    SM --> ST["💾 Storage"]
    SANDBOX --> ST
    WF --> ST

    classDef user fill:#e3f2fd,stroke:#1976d2,color:#0d47a1
    classDef core fill:#fff3e0,stroke:#f57c00,color:#e65100
    classDef cap fill:#f3e5f5,stroke:#7b1fa2,color:#4a148c
    classDef infra fill:#e8f5e9,stroke:#388e3c,color:#1b5e20

    class U,SM user
    class T1,T2,T3 core
    class TOOLS,SKILLS,SUB cap
    class SANDBOX,SCHED,WF,ST infra

In the past, almost all of this was done for you by the vendor. All components were packaged into a software "shell," like Claude Code or Codex.

What you used was essentially a finished product. The internal structure was all hidden behind the scenes, invisible and unnecessary for the average user to see.

The only places an average user could really tinker were things like Skills and MCP. As for the stuff inside the shell, it was all pre-packaged by the vendor—you couldn't change it, and generally wouldn't think to.

An analogy: The old Harness was a fully-furnished apartment, move-in ready with a great experience, but you couldn't knock down a wall or change the layout.

Now, DeepSeek Harness has directly broken all these components down into independent plugins. Developers don't need to modify the Harness's source code to freely choose, replace, or extend any of its capabilities.

Simply put, everything in the Harness is fully customizable by you.

This is also the design principle emphasized by DeepSeek officials: Everything is a plugin.

The official website is here:

https://www.deepseek.com/harness/

DeepSeek Harness actually has only one true core, a kernel called Cordis. This kernel can swap plugins at any time during the Agent's operation while ensuring the Agent's running state doesn't crash.

Simply put, the Cordis kernel is only responsible for loading, unloading, and dependency management of plugins; it doesn't care about anything else.

So, now you should know why DeepSeek named it Harness. Because its positioning is to be a Harness infrastructure system, not merely an AI Coding tool.

Through this infrastructure, or rather, through this everything is a plugin design philosophy, developers worldwide can participate together, plug and play, and create all sorts of plugins to enrich the entire DeepSeek Harness ecosystem. This builds a new foundation for the AI era while simultaneously enabling the Agent's self-evolution.

It's worth mentioning that after being open-sourced, DeepSeek Harness became the fastest-growing open-source project by stars in GitHub's history. As of finishing this article, it has reached over 150k stars on GitHub.

Project address:

https://github.com/deepseek-ai/deepseek-harness

After discussing what it is and its design philosophy, what ordinary users care about most is how to use it. Next, let's go from 0 to 1 and install and set up DeepSeek Harness in a local environment.

Step 1: Environment Preparation

You need to have Node.js installed on your computer first. Open the terminal and type:

node --version

Seeing a version number means it's installed. If not, go to https://nodejs.org/zh-cn/download to download and install the LTS version.

Step 2: Install DSH

Once Node.js is ready, install DeepSeek Harness globally:

npm install -g @deepseek-ai/dsh

After installation, verify it was successful:

dsh --version

Seeing a version number means the installation is complete.

If you just want to try it out temporarily, you can also choose to start it online using the npx command:

npx @deepseek-ai/dsh web

If you are a technical professional and anticipate needing customization later, you can directly choose to install from source code:

git clone https://github.com/deepseek-ai/deepseek-harness.git
cd deepseek-harness
pnpm install
pnpm run build
pnpm dsh web

If you're a layperson and don't understand the above commands or how to run them, just throw these commands to any of your local Agent products (Codex, WorkBuddy) and let it install for you.

Step 3: Start the Service

Enter the following command to start:

dsh web

The service will start locally, and the browser will automatically open http://127.0.0.1:3080. If it doesn't redirect automatically, manually enter this address into your browser's address bar.

Step 4: First-Time Configuration

When you open the interface for the first time, a beta testing disclaimer will pop up. Harness is still iterating rapidly, and interfaces and features may be adjusted at any time. Click "Continue" to enter the main interface.

After clicking continue, the first run will ask you to fill in your DeepSeek API key. Go get one from the official website and remember to top up some money.

If you don't have an API Key, you need to go to the API open platform official website first: https://platform.deepseek.com/ to register and create one.

Copy and paste the created API KEY into DeepSeek Harness, and you can enter the homepage.

Step 5: Select Mode

After configuration, click the "New Session" button on the left to enter the conversation interface.

Create a directory DeepSeekProjects as your workspace:

mkdir DeepSeekHarnessProjects

When creating a new session, you can choose the operating mode. DeepSeek Harness currently has four built-in operating modes, which is also the hardest part for ordinary users to understand and what confuses people the most.

Let's get to the point: These four operating modes are not absolute. You can understand them as a few default preset templates provided by DeepSeek officials. Yes, purely templates, for your convenience right out of the box. You can also customize new modes.

Let me briefly explain the differences between the four built-in modes.

Standard Mode

As the preferred mode for beginners, if you are using it for the first time or don't know much about the underlying mechanisms, just go with Standard Mode without thinking. Standard Mode has complete code Agent capabilities, including file reading and editing, Shell, file search, web search, Skills, planning, goals, background tasks, sub-Agents, and workflows, ready to use out of the box.

PTC Mode

PTC Mode has all the capabilities of Standard Mode, with the main difference being in how tools are presented.

In Standard Mode, the model typically initiates a tool call, gets the result, and then decides the next step. PTC Mode provides the model with a Code Mode SDK, allowing the model to write a TypeScript program that combines multiple tool operations within a single run_code call.

So what's the use of this?

In plain terms, operations that might have required five model round-trips for reading, searching, filtering, parallel calling, and result collation can potentially be compressed into a single program execution. This reduces the number of back-and-forth dialogues between the model and tools, saving tokens.

Of course, it relies on strong code planning capabilities and is harder to debug; beginners shouldn't rush to touch it, just use Standard Mode. It's recommended to switch only after you're familiar with it or encounter a large number of repetitive call scenarios.

Minimal Mode

You can understand Minimal Mode just by its name. This mode only retains the capabilities of a persistent Bash and a file editor.

This mode is mainly used for model benchmarking in a minimal environment.

For example, use it when you want to compare the raw Agent capabilities of two models. But this mode is not suitable for daily development.

Creation Mode

It also has all the capabilities of Standard Mode, but it does one more thing: it can inspect its own running Cordis environment, experiment with plugins in memory, and help you create new Agents and plugins.

Frankly, it allows the Agent to modify itself directly.

For instance, you can ask it to "mold" a brand new Agent:

"Help me create an Agent that is only allowed to read code, cannot modify any files, and is specifically responsible for code auditing."

You can also mold one tailored to your workflow. For example, if you are a tester:

"Help me create a testing assistant Agent that connects to the company's Jira, can read requirements, submit bugs, but can't touch anything else."

But these two are just about "creating other Agents." The truly magical part of Creation Mode is the third play—modifying itself.

For example, if you ask it to run an API stress test, it will first check if it has a stress testing plugin on itself. Finding none, it won't reply with "Sorry, I can't do that." Instead, it will write a stress testing plugin on the spot, attach it to itself, and then finish the job.

The whole process, you only said one sentence.

To use an analogy, it's like an Agent realizing it doesn't have a wrench, so it builds a wrench on the spot, plugs it into its hand, and then continues working with that wrench.

This is the most special part of DeepSeek Harness and the best embodiment of the Cordis kernel.

Step 6: Start a Conversation

After selecting the mode, enter your request in the input box, and you can start conversing with DeepSeek's Agent.

For example, I asked it: Do you think the DeepSeek Harness design philosophy is good? If so, what's good about it? You can only state three points.

Additionally, DeepSeek Harness provides a Trace feature, which allows you to view each run by source. For developers, this means observability, auditability, and reproducibility, making it very suitable for research.

Step 7: Using Plugins

As mentioned earlier, the most important design principle of DeepSeek Harness is: Everything is a plugin.

Plugins offer two ways to play:

One is to create your own plugins. The benefit is that it helps you understand how Harness registers tool plugins, suitable for learning purposes.
The other is to install plugins written by others, ready to use, suitable for daily use.

Since this tutorial is mainly aimed at beginners, I won't cover how to create your own plugins here, though it's actually not difficult.

For daily use, if some features of DeepSeek Harness don't meet your needs, you can directly use plugins made by experts.

The official community currently lists many plugins:

https://github.com/topics/dsh-plugin

For example, I picked a plugin for one-click skin changing for DeepSeek Harness: dsh-web-ui

Installation is also very simple, just one command:

dsh plugin --profile web add @linxin666/dsh-web-ui-all

After installation, restart dsh web, open 'Settings > Plugin Configuration' to toggle plugins on/off as needed, or try on skins in the skin center.

There are many more fun plugins available for you to install and experience on your own.

Step 8: Connecting Third-Party Models

As explained above, in DeepSeek Harness, everything is a plugin, and anything can be replaced. Models can certainly be swapped too. If you want to connect other vendors' models in DeepSeek Harness, it's entirely possible.

Here, I'll use GLM-5.3 as an example.

Demonstrating how to connect a third-party model into an existing DeepSeek Harness (DSH for short).

Open the DSH Web interface, click the 'Settings' button in the lower left corner, go to the 'Models' tab, and click 'Add Provider'.

The 'Model Provider' list already has dozens of mainstream domestic and international vendors built-in for free selection. Here, we choose the model provider as zai-coding-cn, then go to the Zhipu Open Platform (https://bigmodel.cn/apikey/platform) to get your API Key and fill it in.

Expand the custom settings and click 'Fetch Available Models'. However, because GLM-5.3 is too new, it won't be automatically pulled by default.

So if you need to use the glm-5.3 model, you need to manually click 'Add Model', enter glm-5.3, and then save.

After configuration, you can select and use the GLM-5.3 model in the dialog box.

Use the same method to continue adding your GPT, Claude, Kimi, and other models.

Final Words

DeepSeek Harness is just a starting point. What's truly worth watching next is whether "everything is a plugin" can form a stable enough development standard for the industry, and whether it will grow into DeepSeek's own Coding product or become the underlying Harness commonly adopted by more Agent products.

Anyway, the other interesting aspects of this Harness are left for everyone to explore.

Note: What has been released so far is only the Developer Preview version of DeepSeek Harness. In plain terms, it's an early test version specifically released for developers. Ordinary users can also try it, but it is not recommended for use as a daily driver.