Original video in Chinese.
Key Takeaway
- Software development is shifting to Agent First: from human-facing (GUI/TUI) to agent-facing, with tools becoming CLI-based, the Skill ecosystem exploding (now over 30,000), and the industry consensus becoming that all capabilities should be directly callable by agents “like a native language.”
- The newtype system is fully CLI-ified: I completely transformed my multi-agent content production system (8 agents + Skills) into a command-line interface, allowing external agents like OpenClaw to call it directly and become a professional content outsourcing team, with a massive boost in productivity.
- Four-layer agent tech stack: CLI (execution interface) + MCP (deterministic tool integration) + Skills (declarative capability encapsulation) + Framework (multi-agent operating system). I’m using this to build a complete closed loop, and independent developers must seize this wave of agents becoming the main force of productivity.
The direction of software development is shifting from being human-facing to being agent-facing.
This is a feeling I’ve been having very strongly lately. So the day before yesterday, I made a decision: to completely CLI-ify the newtype content production system.
newtype is a multi-agent orchestration system, with 8 built-in agents that make up a professional content team. In addition, I added several Skills to strengthen capabilities in analysis, writing, and more.
Previously, this production system was always for people to use. After startup, it would appear in the form of a TUI. When I say CLI-ification, I mean not needing a graphical interface, and instead operating directly through the command line.
For example, entering nt research plus a topic will launch the researcher to do research; when that’s done, the fact-checker will do fact-checking.
I’ve prepared a whole set of commands like this. Obviously, they aren’t for humans to use, but for agents. For example, your OpenClaw can call newtype, which is like hiring a professional content outsourcing team—the results are definitely better than having OpenClaw generate content on its own.
To teach agents how to use newtype’s commands, I specially made a Skill as well. To make installation easier for users, I added an nt init command, which scans the installed agents and then stores the Skill into the corresponding folder according to the specification.
That strong feeling I mentioned at the beginning actually came from OpenClaw. Really, starting with this wave of OpenClaw, “Agent First” has become the consensus of more and more developers.
Let me share two very obvious phenomena with you, and you’ll know why I say this.
First, various tools are starting to become CLI-based.
For example, Firecrawl, Supabase, Stripe, and so on—these very famous tools in the market have all launched CLIs.
Firecrawl did the most extreme version. They released a dual toolkit of CLI and Skill. With a single command, you can complete CLI installation and register the Skill to all detected agents. The method I mentioned earlier was inspired by Firecrawl.
Second, the Skill ecosystem is exploding.
At the beginning of this year, there were already more than 31,000 Skills circulating in the community market. After Vercel launched the skills.sh platform in January, it quickly surpassed 20,000 installs.
This move toward Skill-ification isn’t something only a few teams are doing; it’s a collective industry-wide action, including Anthropic, Vercel, Supabase, Cloudflare, VS Code, and more.
The essence of CLI is to standardize a capability’s execution entry point into the atomic operations of text input and output, so that agents can call it directly like speaking a native language.
The essence of Skills is to package “when to use a certain capability, which tools to call, and what steps to follow to complete the task” into declarative instructions, so that agents can directly learn how to use that capability.
This is what I meant at the beginning: the direction of software development is shifting from being human-facing to being agent-facing.
So this year’s trends and roadmap are very clear. Agents are the main characters. Everything is being built for agents. The entire tech stack includes four layers:
CLI is the execution interface. More and more tools will definitely launch CLI versions. The share of software operated by agents will definitely keep growing.
MCP is the tool integration protocol. Don’t look at how many people are criticizing MCP. But if you’re working on engineering projects rather than knowledge projects, and you need deterministic action execution, then MCP is definitely necessary.
Skills are declarative capability packages. What they provide is non-deterministic knowledge delivery. They’re flexible, lightweight, and easy to create. Ordinary users and self-media creators can get started quickly, which is why they’re so popular.
Framework is the agent operating system, including memory, scheduling, and routing. OpenClaw and the newtype I made both fall into this category.
These four layers are actually a whole. Let me use an analogy to make it easy to understand: Framework is the foundation, Skills are the manual, CLI is the door, and MCP is the pipeline.
This is how I judge it, and this is how I’m doing it:
newtype’s Framework is a multi-agent orchestration system paired with a memory system.
On top of this framework, I add a series of Skills to enhance capabilities in certain areas, which are called by agents as needed.
This Framework and these Skills together form the core of newtype.
As for access, when running locally, if a person is operating it, then it goes through the TUI; if another agent is operating it, then it goes through the CLI.
If I ever want to deploy it in the cloud later, I’ll probably package it as an MCP Server.
If you’re an independent developer, you absolutely need to seize this wave of trends. The share of agents as the main force of productivity will definitely keep growing.
OK, that’s all 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!