Skip to main content
After graduating in computer science, I spent four years working as an analyst in finance. Four years of coding and four years of writing put me at both ends at once: among the people who build tools, and among the people who live in them. After following generation after generation of compute chips, and after many late nights wrestling with disconnected research reports, Excel models and industry data, I noticed something. In real professional work, the rare judgments, the ones that often take only a moment, are separated from the finished result by a mountain of simple but tedious work: collating, formatting, polishing. These are skills any intern is expected to have. Perhaps every industry is still waiting for a genuinely efficient tool. After ChatGPT broke through, I built Dify workflows and agentic search, and designed PaiWork, a vertical workbench for finance that users loved. That hands-on work, and the wider view I gained after joining MiniMax, led me to an architectural conclusion that keeps getting sharper: more than an agent, every industry needs a native environment where agents and people work closely together in one cohesive place. That is where the idea of Mini App came from. In short, a Mini App is a product that:
  1. Installs as a standard plugin, where every capability is pluggable;
  2. Calls the Agent itself and all of the Agent’s external capabilities (MCP / CLI), and shares the full context (projects and conversations) with the Agent;
  3. Exposes all of its own capabilities back to the Agent;
  4. Comes with a complete frontend, with one core mission: letting humans and AI work in the same environment.
I want to unpack the thinking behind Mini App from two angles.

​
1. Why a coding agent makes a poor personal workbench

Put a modern domain expert’s work next to a programmer’s and two things stand out: attention is split in almost the same way, yet the foundations underneath are completely different. Same kind of work, different foundations: programmers and domain experts have the same shape of work in their environments, where context lives and how attention is split, but the foundation agents rely on is completely different A modern domain expert works inside professional software and SaaS tools: EDA for chip engineers, CAD for industrial designers, trading terminals for financial analysts, and Office for nearly everyone. Their context is scattered across every corner of the system. More than half of their attention may go to understanding the project and reading the wider environment; the actual “writing” takes less than half. The same is now true of programmers who already rely heavily on coding agents. But the same structure on top does not mean the same foundation underneath. Today’s office agents are fixated on one-shot delivery, which is a cardinal sin for productivity tools. Even in professional software engineering, very few projects can be written once and shipped. And code at least has compilers and scripts to pinpoint defects; in most industries, the only checker is a person. AI in the lab can solve a lot of problems. But the real moat in every industry is the loop: ask a question, solve it, then ask the next one. While an expert reads the whole environment and frames the question, the judgments that set the direction live in their feel for the full context. No black-box “one prompt and done” can stand in for that. So we arrived at the first line of the Mini App philosophy:
Real productivity in every industry is still rooted in highly specialized work environments. A plain chat UI cannot replace the precise control and rich views of professional tools. The professional workflow of the future should not ask users to leave their workbench to accommodate AI. It should let professional environments and AI grow together.

​
2. Not an entry point: building the agent into the computer as a system service

In the coding-agent era, the lucky few that became the entry point may already be earning tens of billions of dollars in ARR. But measured against the end goal of “Intelligence with Everyone”, programmers are a small group. An entry-point model built for that small group, one that already shares Bash as a common protocol, simply does not carry over to every other industry. Recent moves across the industry keep proving the point: try to take over everything through a single surface-level agent entry point, and you inevitably break who owns the interaction and how the ecosystem is ordered. That is why, in building Mini App, we passed on the narrow door of grabbing the entry point and chose an architecture that goes deep into the foundation: Agent as OS: the domain expert owns the interaction; Mini App workspaces call both Agent OS and classic system services; Agent OS unifies context and tool protocols, schedules agent tasks and serves inference, and runs on classic services such as memory management, thread scheduling and sockets; CPU / GPU sits at the bottom If compute is electricity, the agent should be the operating system. A good harness should not be built as an entry point. It should run quietly underneath, like kernel scheduling or a file system: taking over the context and tool protocols scattered across the system, and turning AI into an atomic capability that any professional environment can draw on. Once the agent is built in as a system service, the modern computer finally becomes a cohesive professional workspace, where plugins are software and plugins are apps. Agent plugins let you reshape the agent; Mini Apps let you build applications on top of it. The host takes care of everything at the system level, and developers only write the interface and the business logic. Going one step further, we got clear on what Mini App sets out to do:
By building professional-grade Mini Apps that grow on top of Agent OS, we aim to connect the specialized interfaces of every vertical with the AI-native compute underneath, and reshape how every industry works.
A few fun Mini Apps