Original video in Chinese.
Key Takeaway
- Agent as the new content: In the AI era, content is shifting from an information carrier to a service carrier. Agents package knowledge, methodologies, and tools, and can directly help users get work done.
- newtype Profile system: This open-source plugin is built on OpenCode + oh-my-opencode. It orchestrates a multi-Agent content creation team, with a Chief main Agent breaking down and assigning work to Deputies, who research, analyze, and generate the Newsletter in sequence, avoiding broad refusals.
- Advantages of open-source customization: Claude Code cannot support deep customization; open-source frameworks allow full modification. I migrated my entire system over, iterate on it frequently, and plan to inject my personal style into it.
In the AI era, what content creators produce should no longer be limited to articles and videos. It should also include Agents.
I think Agents are the new content.
Think about it: you read an article or watch a video, and that’s it—it’s just information.
But Agents are different. They contain my knowledge, my methodologies, and the tools I’ve tuned. Once you have them, they can help you get work done.
That’s what I mean: content has shifted from a carrier of information to a carrier of services.
I believe this is definitely the trend. And at the moment, not many people in the world are doing it this way, though Dan Koe might count as one.
Based on that judgment, the first product I made this year was a multi-Agent orchestration plugin called newtype Profile. Once you install it on OpenCode, you have an entire AI content creation team.
For example, at the end of last month, Nvidia initiated an acquisition of Groq. I asked the Agents to research the event and generate a Newsletter.
After receiving the request, the main Agent called Chief determined that this was an execution task rather than a discussion task, so it delegated the work to its deputy, Deputy.
However, because the request was too broad, Deputy firmly rejected it. So Chief broke the task down again, started from the most basic level, and conducted deep research in sequence.
At this point, Deputy began assigning specialized Agents. For this task, it had the specialized Agents do two things: first, check whether there were any relevant documents in the local knowledge base; second, search the web for more information.
Once all the materials had been collected, Deputy would form an execution summary and send it back to Chief. After Chief deemed it OK, it would issue instructions again and begin writing the Newsletter.
In the end, an article that Chief considered very high quality was completed and saved in the local folder.
I posted this article on my WeChat public account, newtype AI. If you’re interested, go take a look. I think it was indeed pretty well written.
From the example above, you can see that these Agents form a three-layer architecture with very clear divisions of labor:
Chief is the thinker.
Deputy is the executor. It is specifically used to coordinate specialized Agents.
The specialized Agents are all experts, and they cooperate with one another to complete the task.
This is a system more powerful than Claude Code.
I know some people will definitely say: Claude Code also has Sub-agents.
But honestly, that only counts as entry-level multi-Agent orchestration, and its performance is very limited.
So what makes my architecture good?
First, cost optimization.
Chief only makes decisions, so I use the Opus 4.5 model for it, the most powerful and also the most expensive.
Deputy is responsible for coordination. So I use the Sonnet 4.5 model for it, which is enough.
The other sub-Agents need huge context windows for content creation, so I equip them all with Gemini models, though some use Pro and some use Flash.
You can see that the model allocation is completely customized, so the cost is definitely optimal.
Second, intelligent collaboration.
After Deputy receives the task, it summarizes and filters it before sending it back to Chief. This prevents Chief’s context from blowing up.
In addition, the Agents also have an Artifact system for sharing structured data. Once the previous Agent finishes its work, it can pass the necessary data in a clear structure to the next Agent.
Third, quality control.
I designed a multi-dimensional quality scoring system for each sub-Agent. After each sub-Agent completes a task, it scores itself and evaluates whether the task has really been completed well enough to be handed off, or whether it needs rework.
For example, Researcher is the intelligence officer, so its three evaluation dimensions are coverage, source quality, and relevance.
Writer is the writer, so its three evaluation dimensions are structure, clarity, and being evidence-based.
When the score is above 0.8, it can be handed to the next Agent for processing. Between 0.5 and 0.79, it optimizes on its own. Below 0.5, it has to be redone, up to two times.
With this multi-dimensional quality scoring system in place, the output quality of the entire pipeline is guaranteed, and each sub-Agent also knows what it should revise.
See, that is what “multi-Agent orchestration” really means.
Don’t think that multi-Agent orchestration just means running a few sub-Agents in the background—you have to consider a lot of things, including the ones I just introduced.
And these are all things Claude Code doesn’t have and can’t do. To achieve this level of customization, it has to be an open-source project.
Because OpenCode is open source, there is the oh-my-opencode plugin. It brings a framework for multi-Agent programming.
Because oh-my-opencode is open source, I’m able to make comprehensive modifications on top of its framework, which is how newtype Profile came into being.
And my plugin is also open source; it has already been released on GitHub. Everyone can use it freely, and you can also fork it and modify it however you like.
I’ve already migrated my entire AI system over here and use it at high frequency every day. And I iterate on it as I use it. This plugin was entirely made inside OpenCode using the oh-my-opencode plugin.
Next, I plan to add more stylized elements so that the Agents have a bit of my vibe.
OK, that’s it for this episode. If you want to understand AI, want to become a super individual, and want to find like-minded people, come join our newtype community. See you next time!