DeepSeek Harness Is a Developer Powerhouse, but Casual Users Should Wait

Recently, the biggest event in the domestic AI circle should be DeepSeek Harness.

I also followed up and seriously experienced it.

After using it, I can sum it up in one sentence: It's really powerful, but the current version is not yet suitable for ordinary users.

01 Where it truly excels

Just as advertised on the official website, the truly powerful part lies in: Everything is a Plugin

Let me translate this sentence and progressively deepen its meaning.

1. You can customize your own Agent.
2. You can customize your own Agent within an Agent.
3. You can customize your own Agent through an Agent within an Agent.
4. An Agent can customize itself.
5. Agent self-evolution.

With this progression, the ultimate intention becomes very clear. This is also the true ambition of DeepSeek Harness.

Besides the advanced architecture that exceeded industry expectations, there is another design I particularly like. It's called Trajectory.

Although other mainstream Agents also have execution processes, they are relatively coarse. The dialog box often only shows polished summaries, without seeing the details.

In each conversation round, what exactly the Agent called, where it got stuck, how many tokens it burned, what the context was, and the specific tool calling process.

In DeepSeek Harness, Trajectory lays out all these raw processes, distinguishing them on a timeline with different colored blocks. Each process can be clicked to see further detail information.

It's a bit like opening the developer tools in a Chrome browser to view HTTPS requests. It also feels a bit like a packet capture tool.

02 For a missing feature, let it find and install a plugin itself

I asked it to sort out its own architecture, and it output a mermaid script, which couldn't be previewed directly.

Clever me directly asked the Agent to find a mermaid plugin itself and install it for me.

It lived up to expectations, found 3 for me, recommended one, and after I confirmed, installed it for me. Immediately after, it needed to restart the web service.

After the restart, the architecture diagram in mermaid format displayed normally.

For developers, this is very normal. For ordinary users, this step has a cost.

03 I asked it to write a plugin for itself

After all, it's so hot right now, not casually writing a plugin would leave me feeling empty inside.

But what to write? It's mainly for experience and practice. Clever me thought again, let the Agent think of something for me.

Sure enough, it praised me. Said it was a "good idea." And then busily finished its research, giving me a result. I looked, good heavens, Chinese people are really competitive. Only 3 days since release, and there are already so many plugins.

At the same time, it suggested 3 directions. And specifically recommended the first one. I have to say, I had actually encountered the first pain point before.

Then I asked it to make a clipboard history manager. Monitor the system clipboard, save the history, and allow searching and copying in the sidebar.

We reached a tacit agreement, and then it busily went off to work again.

Not long after, the plugin was quickly produced. Named Clipboard History Manager, status was pending approval. It needed to be approved.

I was just about to give a thumbs up, but unexpectedly, the trouble began.

04 Not smart enough, over an hour, back and forth

It verified the monitoring and persistence itself. But the UI was no good.

The clipboard button wasn't visible at the bottom of the sidebar. Refreshed, still not there. Later, the button did appear, but only partially visible.

I wanted to throw a screenshot at it. The interface directly prompted: The current model does not support images.

I could only use text descriptions, but this problem was revised many times back and forth, with no effect.

Finally, it thought of a solution itself. It moved the clipboard entry from the bottom left corner to the top right toolbar.

What a sly old fox. Alright, let's use it for now.

Later on, I asked it to change this dynamic plugin into a static plugin. It needed to restart the web service once more.

After the change, the delete function had problems again. This bug also went back and forth, revised several times, and finally got fixed.

Finally, it was published according to the community dsh plugin specification. This step went quite smoothly.

After installation, clicking the clipboard in the top right corner, it looks like this when opened.

In total, it took over an hour. Honestly, I was dissatisfied with the time the Agent wasted in the middle. It felt not very smart, going back and forth several times. An average user would have lost their patience long ago.

05 Tokens are really cheap, but prices are going up soon

Overall experience, the cost was also really cheap. The cache hit rate was really high.

Time was spent, but not much money.

The whole process consumed 2.6 yuan. The cache hit rate in the session could reach 98%, which is really high. Official notice says prices are going up soon.

06 The ecosystem is already fiercely competitive

What truly surprised me was the speed of the community.

Just 4 days later, the community already had over 6,600 plugins, and the DeepSeek Harness GitHub repository had reached 147K stars.

Chinese people are really competitive. The technical enthusiasm is also really high, the power of the ecosystem is too strong.

I have to admire, DeepSeek's move is quite fierce.

07 Currently not suitable for ordinary users

After installation, open the local address. The current official UI is in web form. Version number 0.1. The popup window writes very bluntly: For Harness developers, core plugins and APIs will iterate rapidly.

This already clarifies the user profile. This version is not a product prepared for ordinary chat users.

The interface looks familiar, similar to the DeepSeek web version. Just that conversations have become local workspaces. There's no vertical three-column layout like Codex yet.

Overall feeling, it's still a rough shell currently.

Additionally, regarding multimodality, I think it's very practical, I don't know why DeepSeek has been slow to support it. For the interface issues mentioned above, three sentences from you are not as good as throwing one picture. Now you can only rely on text descriptions.

Many scenarios require throwing in a picture, or even a video, combined with text. This is very common in work, and Codex supports this quite well.

Currently, there are no scheduled tasks, which are actually quite practical too.

The built-in plugin display is also too simple. In the plugin list, there's just a bare name, and for some, you can't even tell what they specifically do.

Of course, after all, this is only the earliest version, and the plugin community is developing fast enough. These current product experience issues are not problems as time goes on.

I believe that as the product capabilities become richer, making it more convenient for ordinary users won't take too long.