Original video in Chinese.
Key Takeaway
- Multi-Agent orchestration trend: this year, AI is shifting from single-Agent execution to managing 10-12 Agents working in parallel (like a combine harvester). Customization determines the upper bound and is becoming the core of productivity.
- OpenCode pipeline: + Oh My OpenCode plugin enables fully automated Agent collaboration. The main Agent analyzes and breaks down tasks, while sub-Agents execute in the background. Config files/Forks let you customize freely.
- Smart examples and plugins: when it judges a “discussion” mode, it dispatches researcher/fact-checker sub-Agents to investigate and verify, providing context rather than a simple true/false answer; my newtype Profile plugin optimizes content creation, with productivity higher than Claude Code, and is shared on GitHub/in the community.
The big trend in AI this year is still Agent, but it’s not quite the same kind of Agent as last year.
Let me give you an analogy. Last year, we were harvesting corn with a sickle; it was still very primitive. This year, we’ll be driving a combine harvester and becoming farm owners.
It’s not about staring at one Agent writing code, but about managing 10 to 12 Agents working in parallel.
A dozen Agents running in the background while we sip coffee and wait for acceptance. That’s roughly what it looks like.
“Multi-Agent orchestration” will definitely be the biggest trend in AI this year.
Under this trend, I believe there is one more thing that will definitely happen, and that is customization of these Agents.
Because in production environments, the upper bound of an Agent is not only related to the model’s performance; another very important deciding factor is whether it can run according to the way you work.
That’s why OpenCode and its plugin Oh My OpenCode have become so popular recently.
This combination provides a fully automated Agent pipeline. After you give the main Agent a requirement, it automatically analyzes and breaks it down, hands it to different sub-Agents to execute, and runs a bunch of tasks in the background.
And this thing can be changed however you want. You can make small tweaks through config files, or you can Fork the project and redesign it.
Let me give you an analogy.
Claude Code is like a well-equipped mass-produced car. It’s very comfortable, very easy to use, and it drives right off the lot. It also has some “modification” options.
OpenCode, by contrast, is a full-on modification platform. It gives you a very solid chassis, and you can modify it into the racing car that suits you best.
With OpenCode and the Oh My OpenCode plugin, everything is laid out for you—from model selection to Agent configuration and even creation. As long as you have the hands-on ability, you can completely turn it into the AI operating system that suits you best.
For example, I built eight Agents specifically redesigned for content creation scenarios on top of the Oh My OpenCode plugin, packaged them into a new plugin, and shared it.
In this episode, I’ll show you how to configure it, and how I completely modified the plugin so the AI operating system fully meets my own needs and habits.
Hello everyone, welcome back to my channel. Modestly speaking, I’m one of the few creators in China who can explain both the Why and the How of AI clearly. What I provide is worth far more than tutorials. Remember to follow. If you want to connect with me, come to our newtype community. This community has been running for more than 600 days, and over 2,000 friends have paid to join.
If you are in China, you can join through Knowledge Planet. If you are overseas, you can join through Substack. My first course, daily newsletter, and exclusive videos are all available in the community.
Back to today’s topic: OpenCode customization.
Before you start configuring, you need to understand two concepts: “user-level configuration” and “project-level configuration.”
After you install OpenCode, there will be an opencode folder under the .config folder. This is the “user-level configuration” folder. In other words, unless you specifically set something in a project, the default will follow this configuration.
And in each project, you can also create an .opencode folder in the project root directory. This is the “project-level configuration” folder. Its priority is higher than the “user-level configuration” I just mentioned. So if you want to make separate settings for a specific project, you can configure it here.
In both of these folders, there will be these two documents: one called opencode.json, and one called oh-my-opencode.json. You can tell from the names that one is the configuration itself, and the other is a configuration plugin.
To sum up, you can look at the table I drew. The logic is very simple: there are two layers of configuration, with the project level having higher priority; and within the two layers of configuration, each has two documents, corresponding to the core and the plugin respectively.
Once you understand this logic, you can start doing it yourself. For example, at the user level, I use opencode.json to set which plugin to use; at the project level, I use oh-my-opencode.json to set which model each Agent calls, and mcp.json to set which MCP servers need to be added to this project.
The examples I gave are only a small part of the settings. There are too many places where the core and plugin can be configured. For example, you can use Prompt Append to add system prompts to a certain Agent. For specific configuration methods, you can check the official documentation, or just ask AI directly.
This is the method of customization through config files. But this method still has limitations: the Agents in the Oh My OpenCode plugin are hard-coded. They’re fixed. For example, their names cannot be changed; their base prompts also cannot be changed.
If you, like me, pursue deeper customization, you can Fork the project and change all the Agents inside it.
Compared with developing projects, I usually do more content work. So, as I said earlier, for content creation scenarios, I redesigned the Oh My OpenCode plugin and created eight Agents. They exist entirely for content.
Among them, in addition to automatically assigning tasks to sub-Agents, the main Agent also automatically switches between two modes: discussion and execution. This corresponds to Plan and Build in programming scenarios, but adjusted for content scenarios and changed to automatic switching.
This automatic switching is crucial. When you chat with it about ideas, it becomes your discussion partner and helps you organize your thoughts; when you say “start writing,” it immediately turns into a project manager and starts assigning work. You don’t need to manually switch modes.
For example, I asked AI: “I saw a claim that Nvidia compute cards have a failure rate of around 9%. Where does this conclusion and data come from? Is it real?”
After the main Agent received the request:
First, it determined that what was needed was “discussion” rather than “execution,” so it independently entered “discussion mode”;
Then it determined that research and verification were needed, especially to figure out what the context of this viewpoint was;
Finally, it dispatched two sub-Agents to work in the background—researcher to check the source, and fact-checker to verify. This process called Tavily MCP search. In the end, it gave a conclusion.
What I’m most satisfied with is that it doesn’t tell me whether this viewpoint or data is true or false, but tells me what the context or scenario is, and what caused the discussion in the first place.
This is just a small example. In daily work and content creation, there will be many situations like this, where multiple Agents need to work together to complete a whole collaborative system.
This content plugin is called newtype Profile. I’ve already released it on GitHub, and I’ve also shared it within the community.
At the moment, this configuration is the most comfortable for me to use, and its productivity is even higher than when I used Claude Code before. I suggest everyone give it a try. Remember, after installing OpenCode, you must also install the Oh My OpenCode plugin. If your project is not programming but content creation, you can install my plugin.
OK, that’s all for this episode. If you want to learn AI, want to become a super individual, and want to find like-minded people, come to our newtype community. See you next episode!
Keep Practicing
This article preserves the usage record from that time. When actually configuring it, if you run into issues with service providers, model IDs, or configuration overrides, first check the OpenCode Model Configuration and Troubleshooting Guide; if you want to organize a single operation into a complete task, then also read the OpenCode Workflow Guide.