Speaker A
PSK has released their Deeps Harness agent. And this is perhaps the first tool in a long time that I looked at and thought, "Damn, now this is interesting." Here is the idea. You can customize almost everything: the tools,
Complete guide to DeepSeek Harness agent installation, customization, and usage with plugin and model integration explained step-by-step.
Key Takeaways
- DeepSeek Harness offers a modular and flexible AI agent framework with plugin-based customization.
- Installation is straightforward and requires minimal technical knowledge.
- Users can safely control agent permissions to protect their system.
- Multiple AI models can be integrated and swapped easily without subscriptions.
- The agent can automate complex workflows, reducing manual coding and saving time and money.
What the video covers
- Introduction to DeepSeek Harness agent and its highly customizable plugin-based architecture.
- Step-by-step installation process using a single command and prerequisite software.
- Explanation of agent modes and permissions, recommending Workspace WR for safety.
- Managing plugins: enabling, disabling, and reducing errors by limiting active plugins.
- Agent presets (sets) for different tasks and automatic set creation via creator mode.
- Connecting and managing AI models with API keys, including adding custom providers.
- Demonstration of practical use by refining a website without coding using natural language prompts.
- Agent’s ability to delegate tasks to sub-agents and integrate multiple AI models like Claude and Codex.
- Cost transparency with detailed billing and usage tracking per model and key.
- Advanced plugin creation by the agent itself and sharing custom plugins for team use.
Chapters
- 00:00Introduction to DeepSeek Harness and Customization
- 01:16Installation Process and Setup
- 02:26Agent Modes and Permissions Explained
- 03:41Managing Plugins and Agent Presets
- 05:03Understanding Plugins and Model Integration
- 06:22Adding Custom Providers and Managing API Keys
- 07:24Practical Use Case: Website Refinement Without Coding
- 08:35Agent Delegation and Multi-Agent Collaboration
- 12:15Cost Overview and Final Thoughts
Full Transcript — Download SRT & Markdown
Speaker A
the agent's behavior, even the interface itself—it’s all plugins, separate pieces that can be added or removed. And you don't have to write your own by hand. The agent will write it itself. Well, you can start with a simple prompt in Russian. And today
Speaker A
I’ll show you what that looks like. Take a look at this, this is the agent board. I built it myself, right inside the interface. Three columns: what's in progress, what's done, and what's archived. For each task, you can see
Speaker A
the steps, time, and expenses. And here, under one task, four sub-agents were already running. I don’t have this kind of screen in either ClodCode or Code. But here, I built it in half an hour, and that’s it. It just appeared
Speaker A
in my workspace window on its own. Today, we’ll install, pay, and get to work step-by-step. And I'll be honest about where it’s still clunky, since it’s an early version. And at the end, I'll show you the bill, how much
Speaker A
everything you saw in this video cost. Only practical stuff, no fluff. Let's go. This is their page, and the whole installation is just one command. But you need a program before that.
Speaker A
Download it from the official site, install it like any other, and you'll use it to install other programs. Copy the command from the site, open the terminal, and paste it. Just hit enter.
Speaker A
The agent installs everything itself. And that is the only technical part in the whole video. Everything will open in the browser after installation, like a regular website. And by the way, any computer will work. You don't need anything powerful here. You can connect
Speaker A
any model you want here. We’ll get to that later, and I’ll show you how it’s done. And here is the first screen. On the left are the working folders. Sessions are stored there too.
Speaker A
In the center is the prompt bar, and below it are four buttons. That’s all you need to get started. In reality, it's a folder: the agent works inside it, reads what's there, and only edits that. Mode. There are four of them here
Speaker A
. We really only need two. Standard: the full set, files, command line, belts, and sub-agents. That’s the one we’ll work in. Creator. Well, the creator mode. In it, you build your own plugins right during the conversation.
Speaker A
We’ll come back to it, I’ll show you everything. Minimal and PTC, uh, niche cases. Overall, let me skip those. Editing—slow down here—read-only.
Speaker A
The agent only watches. Workspace WR writes, but no further than the working folder. Full Access. Well, that's your whole computer too. Be careful. It is best to select Workspace WR, and only enable full access when you know exactly what you are doing and why. And
Speaker A
the model. Right now, we have DeepSeek Flash from the provider as a placeholder. We will return to the model in a minute and connect everything. Let's go to the settings.
Speaker A
The first thing you will see here is the same modes and permissions. The only difference is that in chat you change them for the current session, but here it is by default for new ones.
Speaker A
The plugins section. And there are two screens here. The first one contains the settings for the plugins themselves and the command line. You can limit what the agent is allowed to execute at all. Internet search. You paste a
Speaker A
separate key here. Without it, the agent will not search. And the loop is the cycle the agent goes through. Think, take a step, see what happened, think again. And this is where you set how many calls it makes simultaneously. I
Speaker A
have mine set to 10. The second screen shows the full list of plugins already installed. By default, I have 166. And each is enabled or disabled with a simple checkbox. Don't need a plugin?
Speaker A
Turn it off, and the agent won't know about it anymore. Fewer plugins, fewer reasons to make mistakes. And that actually works. And the last section is agent sets. Also known as presets.
Speaker A
These are the same four modes you saw in the chat. A set is, well, the agent's configuration. What tools it has, what instructions it follows, and what it is allowed to do. Any built-in one can be copied and customized to
Speaker A
your needs. One set for analysis, another for drafting. Or you don't have to build them by hand; you can ask DeepSeek Harness to assemble a set for you, and it will build it itself in creator mode, of course. That is truly
Speaker A
cool. And we will return to that later. And the section we came here for: models. Nothing will work without a model. And this is where you paste the key. By default, the provider is DeepSeek. And by the way, there is an
Speaker A
"Add your provider" button below. And here is one thing without which half the point is lost. The developers have it right on the main page: the formula "Agent = Model + Harness." The model, as they write, is the soul of the agent,
Speaker A
but by itself, it doesn't read your files, doesn't run commands, and doesn't remember yesterday's conversation. Everything else is there so the agent knows where it is, knows how to use tools, and doesn't fall apart during real work. And right there,
Speaker A
as a header across the whole screen, is "Plugin." It's everything: tools, skills, session, storage, interface, and the model is just as interchangeable a plugin as the rest.
Speaker A
That is why today I would choose an agent, not a model. The model here is a part that you swap out; if you don't like it, you put in another one. Need Claude or Gemini? Try them the same day
Speaker A
and worry about the money later. There’s no subscription here, just a key. You pay for what you use. I think that’s even more convenient right now, given the recent blocks. I’m in Russia, so I’ll add my own provider.
Speaker A
This is Polza AI, a model aggregator. And there are really a lot of models here. Well, first, of course, we register, and then we top up the balance with a card or via SBP. A few hundred rubles is enough to start.
Speaker A
After that, it just deducts as you go. From the models, I choose DeepSeek V3, the new version. It was updated recently too. It’s inexpensive and works well for my tasks. There are other models nearby as well. I’d definitely note the DeepSeek V3 Pro.
Speaker A
We’ll take a look at it today too. And there are American models available as well. Now for the key. I’m just creating it. You set a limit here, for the amount and the period. I set it to 200 rubles. For a week, so I don’t
Speaker A
accidentally go to zero. You can see your expenses later in your account. And, by the way, for each key and each model separately. Very convenient. How to connect it. There is full documentation. I’ll show you everything right now. There are
Speaker A
ready-made instructions for different agents. Like Aider, Cursor, Cline, and many others, though it connects to everything. Here is the address where Polza accepts requests. I copy it, paste it here, and add /v1. Then the key we already created and copied. I click
Speaker A
create. Although it asks to add at least one model. Opening the list, there’s the entire Polza catalog.
Speaker A
Everything is already checked. I can choose or just leave it and add everything. Let’s just add it all.
Speaker A
The link to the service is in the description. And that’s it, our key is in place. We update the models and can get to work. Alright, enough settings, let’s actually get to work.
Speaker A
We’ll take the site I built in the last video on Claude without a single line of code. Now we will refine it, also without a single line of code.
Speaker A
I’ll ask in plain language, and the agent does the rest itself, just like last time, only now it finds the issues too. What needs fixing. I’m showing this on a website because you can see the results with your own eyes. It was
Speaker A
broken and now it works. And honestly, almost anyone can build anything using text prompts today. And it truly works for personal use. But I wouldn’t hand this over to a client. It looks like a storefront on the outside, but I’ll
Speaker A
show you what’s inside. And a quick disclaimer for those w
Speaker A
mechanics are exactly the same. The agent reads everything in the folder, identifies issues, and organizes them into tasks. Don't look at the code; look at the workflow. I’ve already downloaded the website project, put everything in a folder, gave it to the
Speaker A
agent, and asked it to create tasks for everything it finds. And it got to work . Here’s the result: 18 tasks, and it didn't just dump them in the chat. It created a file right in the project for
Speaker A
each one, detailing what’s broken and how to fix it. And it prioritized them by importance on its own. Look at the top two. First, the site has a "10%off the second item" promotion. It’s displayed on the screen, but it’s not
Speaker A
included in the calculation; items hit the cart at full price. Which is absolutely wrong. The customer was shown one figure, but charged another.
Speaker A
The second one is the "Place Order" button. It doesn't collect or verify anything. The required fields are only required in appearance. I clicked it, the cart cleared, and a pop-up appeared . "Thank you for your order," but it
Speaker A
didn't actually go anywhere. The agent wrote that the discount wasn't working. I checked it: opened the site, found the promotion on the product page. The discounted price is crossed out, with the full price next to it. I add it to
Speaker A
the cart and go to checkout. Yep, full price. Everything is just as it said; it was right. By the way, for the site to even open in the browser, the agent set up a local server itself. I just
Speaker A
asked it to do that, too. The first thing this run taught me: You have to read everything. Some of the findings weren't bugs, just hidden scrolls, empty links, or notes about future filters. That’s how it was intended.
Speaker A
The agent doesn't know your intentions; it only sees the code. We weed those out immediately, otherwise, you'll be fixing things that aren't broken. Well, since this bug is real, let's fix it right away. I ask it to fix only this
Speaker A
task and not touch the rest. And it just fixes it, checks the code for errors, verifies that the math holds up with different quantities, and ensures the site opens. I open the promo item; now it's marked with the discount, and
Speaker A
a new line appears in the order. " Discount: -10%." Everything works; it’s logical. And here is the second thing. Look at the total. But they don't add up on the screen. The total itself is correct; there are no
Speaker A
questions there. The discount is included in it, but it shows the "Items " line above it already discounted. So it turns out that 22,000 + 1,000-400 = 22,000. It calculated correctly but displayed it wrong. I caught it with my
Speaker A
eyes in 10 seconds. I pointed it out, and we fixed everything. Take these rules away with you. You need to verify not only what the agent finds, but also what it fixes. Everyone checks the first part, but almost no one checks
Speaker A
the second. That’s how you can keep living with your projects. Once a week, we run agents through the folders, get a list, manually filter out the extra stuff, and fix them one by one. It's just normal work, only the list isn't
Speaker A
compiled by a human. And it's often actually correct. Friends, if this breakdown is useful so far, hit like and write a few words in the comments.
Speaker A
I answer everything myself too. And I don't ask for anything else. It really helps the project a lot right now. And now, for the reason why all of this was started. We have bugs. Fixing them one by one is boring. So we do this: one
Speaker A
session becomes the manager, and under it, four separate ones are launched, one for each bug. Look at the top panel . The sessions appeared on their own and are labeled with task numbers. No one created them by hand. Each one
Speaker A
fixes only its own bug. And each one is in a separate project folder. The tasks are different, but the files are shared . If they access the same file at the same time, they'll overwrite each other's work and it'll be a mess. But
Speaker A
this way, each has its own copy, they physically don't overlap, and in the end, everything is merged back into one project. Moreover, before launching, it verified that all four copies were identical to the original. Here is its plan. Right on the screen, the
Speaker A
checkboxes tick themselves, and the manager doesn't fix anything itself. It watches, waits for the others to finish , and then checks, goes into each result, verifies that the edits don't conflict, and only then merges everything into one file. It caught one
Speaker A
inconsistency between the tasks by itself. I didn't even know about it, by the way. This is what makes the design functional, rather than just launching four agents, where each reports in a standard form what it changed, in which
Speaker A
file and line, how to test it, and what the side effects might be. Well, basically, a standard report. So, it's an agent that launches agents and checks up on them, all in one window.
Speaker A
Well, since there are many agents now, the question arises: what are they actually doing in there? And here is a feature I haven't seen in other agents.
Speaker A
It's called trajectory. A complete log of what the model saw, every tool call, what was processed, and what was returned. You can search through it and replay specific parts of the events.
Speaker A
And now the most interesting part— branching. Here is the message after which the agent took a wrong turn.
Speaker A
Every message has a branching button. I click it and get a separate session with the same history up to that point.
Speaker A
From there, I lead it differently. The original remains intact, and both are visible side-by-side. Why would you need this if an agent hangs or burns three times more money than usual? It's visible here. Exactly on what, and not
Speaker A
based on feelings. And most importantly , it saves time and money. You aren't redoing the half of the work that already turned out well. You've seen the result. Now, here is how you will fine-tune it for yourself. The plugin
Speaker A
store connects with one command, and there are over 2,000 inside. And these aren't amateur hacks among the most popular ones. There's one made by VS Code, installed with a couple of clicks . By the way, this is just one of the
Speaker A
plugin stores you can install. The point here is not to look at the number , as they are counted differently, but at the fact that people have already gathered around the tool in a few weeks , writing and uploading plugins. Many
Speaker A
you can simply take and use. A product with this kind of momentum has a much better chance of surviving. Remember the board the video started with?
Speaker A
Let’s build it right now on camera. I’m opening a new session, not in the site folder, but in an empty folder.
Speaker A
The plugin is written for the tool, not the site. And here I change the model from the junior one to DeepSeek Q4 Pro.
Speaker A
All day before this, I was working on the junior one, the Flash version. And all the site edits cost about 30 rubles . The task now is difficult, so I’m taking the Pro version. And we’ll calculate how much it cost me at the
Speaker A
end. I turn on creator mode and write in plain language. Not code, but a request, a tab where all my sessions and their status are visible. And here is what happens next. The agent digs into its own documentation, reads how
Speaker A
its plugins are structured, writes a new one, and asks me for confirmation to activate it. It registers it within itself and restarts the interface. I just refresh the browser tab; the tab stays put. I click it, and there it is:
Speaker A
our plugin, our board. Three columns: In Progress, Done, and Archive. At the top, three counters. How many tasks are currently in progress? How many tokens were used in total, and how many are waiting for my response here? That’s
Speaker A
the last counter. And it turned out to be the most useful one. You can immediately see where the agent got stuck and is waiting for me. The card shows the steps. Moves, tokens, time, and context usage. Below, the model
Speaker A
it’s running on is listed, along with sub-agents if the task called for them. I hit run, the card moves to "In Progress," and the green counter at the top changes. There’s no need to refresh the page. Let me say out loud
Speaker A
what just happened. The program finished writing itself. With a simple request in Russian, in half an hour, it didn't just generate code for me to paste somewhere. It pulled up its own documentation, figured out its internal structure, and wrote a plugin with its
Speaker A
own logic. It connected everything itself, turned it on, and it started working immediately. I only had to give one confirmation. Claude Code and Cursor are also agents, and you can customize them too. There is a huge selection available. You can choose a
Speaker A
model, even a different one, tweak the parameters, toggle features on or off, and build your own agentic operating system. But you're only adjusting what you’ve been given. Well, essentially.
Speaker A
Whereas here, you’re changing the tool itself. Now, look at what a plugin actually is. In fact, here it is: Agent Bard version 3. Status: running. Next to it is a list of versions—first, second, third—these are its own attempts until a working one was built.
Speaker A
You can revert to any of them. You can disable a plugin with a checkbox or delete it while the program is running, without restarting anything. There. And this is what agentic systems are: not one big program, but a collection of
Speaker A
separate pieces. Each one starts, stops , and updates on its own. That’s why the agent was able to write a piece for you on the fly. Everything else remained untouched. And the dashboard is just my case. It could be anything
Speaker A
instead. Your own CRM, a team task board, or a dashboard for employee agents. It’s assembled just like a plugin inside the agent. I can simply publish mine, and you can install the same one with a single click. That is
Speaker A
the reason why I made this video. It’s not because this is a cheap model. There are plenty of cheap models without this. It’s because it writes code that isn't there, tailored specifically to your task. And this plugin isn't a sandbox; it lives inside
Speaker A
your workspace with full permissions. You can safely write your own, but third-party ones are trickier. A third-party plugin isn't just access to a file, but to the agent. And the agent has your API keys and your projects. If
Speaker A
you grant full computer access, the chain is: third-party code, agent, keys , machine. Compromise one, and well, everything is open. By the way, good news: the plugin won't turn on without you. It asks for confirmation on every new system. Just like it did for me.
Speaker A
And there are two checkboxes: only this version, or all subsequent ones too. Don't check the second one; updates will come, and you won't know about them. So, what do I do? I only install things with a visible author, plenty of
Speaker A
stars, and open source code. Unknown ones go into an empty folder first, where there’s nothing valuable, with permissions restricted to that folder.
Speaker A
You can also invite Claude or Codex to check all of this for you. One more thing that works: this agent can delegate work to other agents. Watch, I’m enabling the Claude Code plus Codex agent mode. And now, from here, I
Speaker A
can call Claude Code and Codex, which are running on this same computer. I ping them, and both respond. Next, you can delegate any task. You can delegate the task to any of them directly from here. And this is what it’s actually
Speaker A
for. If you have paid access, Claude, Code or Codex, this agent can steer them, and they can do the work using your already paid access. So, you take the convenient interface from here, but the brains from there. And from this,
Speaker A
here are the rules I would separate out . I wouldn't choose one tool forever.
Speaker A
Choose the one that will be the dispatcher. There is one dispatcher, but any number of executors. And they can be swapped out painlessly. Once you start thinking in this logic, the debate over which is better—Claude, Code, or Codex—becomes much less
Speaker A
important. Now, honestly, about where it still hurts. The tool is fresh, the version is early; the developers themselves write that much will change later and something might break. This isn't a complaint; it's just the product's stage of development.
Speaker A
Everything you saw in this video is 12 sessions, almost 2 hours of model work, and 19 million tokens. I don't count money for it. In my usual work, I spend , well, about 300 rubles a day on Flash
Speaker A
. And you only pay for the days when you actually worked; if you don't sit at the computer, the bill stays still.
Speaker A
You need a regular, Russian card to pay on Polza. And here is the observation for which I calculated all this. I ran the plugin build on the top-tier, Pro model for 147 rubles. for one task.
Speaker A
Everything else was on the smaller one, Flash. And there were just as many tokens, for 30 rubles. The smaller one is really good. It parsed someone else's code, found bugs, fixed them, and worked in four parallel sessions.
Speaker A
Once it didn't think it through—the totals in the cart, as you saw—but it calculated correctly; run routine tasks on it, you won't notice the difference.
Speaker A
Use the top one when you need an architectural task designed from scratch. To understand an unfamiliar system, or, for example, a plugin build as a whole, I could have used a smaller one, honestly, but I was afraid it
Speaker A
wouldn't work. But 147 rubles. is actually very inexpensive. And if you take only one thought from this whole video, I would leave you with this one.
Speaker A
I wouldn't choose a model, but an agent . Models change every month and catch up with each other. And an agent is everything that surrounds the model.
Speaker A
And that is what decides whether you are working or just playing around. To spend half an hour writing your own plugin for an agent and get a screen no one else has. I haven't encountered anything like this before. Yes, there
Speaker A
are agent systems like Agentix—I have one too—but this is still different. You need to understand that. If you want to try it, start small: install it , give it your work folder, and ask it to figure out what's wrong there. Then
Speaker A
you'll understand for yourself if it's for you or not. And I really ask you to leave a detailed comment, write what was useful, where you still have questions, and which case to analyze next. This helps YouTube's algorithms a
Speaker A
lot, and I reply to you personally. Often, it is your question that becomes the next video. And subscribe so you don't miss the upcoming analyses. Bye everyone. M.
Topics:DeepSeek HarnessAI agentplugin systemagent customizationAI modelsinstallation guidetask automationsub-agentsmodel integrationcost management
