First Look at Verdent, an Autonomous Coding Agent From China

I got onto the recent early access beta for Verdent, the new AI coding tool from TikTok’s former head of algorithms. Described as “an autonomous coding agent that takes care of the boring parts,” there are two flavors: Verdent for VS Code and Verdent Deck. This review looks at the second one, as it is effectively an app interpretation of an agentic Command Line Interface (CLI).

As is common in my reviews, I’ll test it with my Rails project. But I’m not testing the abilities of the LLM engine(s), per se — I’m looking at how the tool workflow executes the tasks for the user. This will be simpler to follow if you are familiar with MVC applications.

Model costs were free on the beta; and as this implies some throttling, I won’t comment much on the speed either way. But everything seemed to work well.

Verdent points out that “Multiple Verdent agents work simultaneously on different components.” That is one advantage of running your own app platform.

Both Kiro (an AI coding tool from AWS) and Verdent look at software development through workflow orchestration — in other words, coding is just one part of the process. In short, this isn’t an assistant to a lone developer, but an emulation of an end-to-end software team.

Verdent Deck

The Verdent Deck app is currently only available for macOS, which might reflect the size of the team — but I assume Linux and Windows will follow after release. After a short verification dance, I was logged in.

👁 Image

The interface is nice and Mac-friendly, feeling more like a corporate website front page than a development tool.

I used the /init instruction to generate the Agents.md file (though this should probably be Verdent.md, unless they are sure to be integrating with other agentic command files). I like the help tip as I typed:

👁 Image

When complete, it gave a nice rundown (left side) of what it had done. I also selected the diff (right side), which in this case is just the Agents.md file contents:

👁 Image

The DiffLens option added more information, in the form of “why we did what we just did” as opposed to “what is this.” I would say the team should focus first on what developers need, and leave compliance links for later.

I noted that the command summary correctly caught the “bin/dev” command that starts Rails and also reloads CSS. Other start methods won’t necessarily do this. The notes about the domain models were also good summaries.

An interesting extra was a mermaid diagram included in the Agents.md file:

👁 Image

If satisfied, you choose to commit this work when done.

There is a little dial to control effort and whether to produce a plan or not. I chose to turn the plan option on and increase the reasoning. This makes it behave somewhat similarly to Kiro later on:

👁 Image

Verdent comes with both a file editor view and a terminal. The icons down the right side are also for the Diff view, and logs:

👁 Image

So let’s do some useful work. At the moment, my list of conversation lines in my Rails app for organising my game’s conversations isn’t organised very well:

👁 Image

Although you probably wouldn’t access it this way, it would be better for this list view to have the conversation name as a heading and for the conversation lines to appear under. So that’s what I’m going to ask Verdent to do.

This is a slightly more nebulous request for the LLM than I usually give, but it clearly understands my model fairly well. We’ll pull up the file, refer to it in the request and see how it does.

I think the right query here is something like “please organise the index view of conversation lines in (file reference) so it lists the conversation lines within their conversation names to improve spatial organisation.” I did this below, and you can see the index file referenced with the “@” command:

👁 Image

Again, this is an advantage of fully controlling your platform’s UI.

Here is the plan it came up with. I think the task has taken on the name “init,” as this was the first command given in the session:

👁 Image

The LLM realised it needed a bit more help from the controller, and recognises that the conversation name is now redundant in the partial view.

Let’s take a quick look at the benefits it lists:

👁 Image

This agrees with my direction, and actually states it in more compelling English. In fact, this would be useful in an agile task description. Note that it asks permission, and we know it needs to change three files.

Verdent methodically works through the plan, crossing out tasks as it goes along, so you see where it is. It does not update the file view in real time.

Once we are done, we get a nice summary (”work completed”) and diff view if needed:

👁 Image

The results are precisely as expected:

👁 Image

We could go on to do more organisational work, now that we have confidence in the system. But this is fine for now.

Lastly, let’s look at the all-important permissions area (Verdent settings):

👁 Image

I think the reference to “allow rules” in the top line is talking about user rules. I’m not sure where I can set roaming rules, but I will assume that it doesn’t work outside of the project directory.

After updating (I was in version 0.6.3 in the short closed beta), I was able to recapture my previous conversation by task (although it was hidden in the sidebar) — still named ‘Init’, which is editable. I noted that support for MCP is available too.

Conclusion

So obviously, creating an app focused on agentic work is a very valid direction. Verdent Deck is currently Mac only, which does point to the usual problem: to produce a Linux and Windows version will need system-specific dev talent. (You can still use the Verdent plugin for VS Code, however.)

Some of the infrastructure behind the scenes is not really visible from the UI, which is why you should probably read The New Stack’s interview with CEO Zhijie Chen to get the fuller picture. One thing he mentions that is both a truth and a subtle warning: the new AI tooling is purpose-built and not an example of “human tools translated for AI.”