Opinion: The Dominance of a Few General Agents in the Plugin Ecosystem Determines Openness
The field of general agents is evolving into a "winner-takes-all" structure, where truly strong general agents will not only encroach on the space of vertical agents but also squeeze the survival space of small and medium general agents, ultimately leaving only a few players with top models and resources.
The current popularity of agents largely stems from the rich plugin ecosystem built around them (such as Skill + MCP), which can integrate model capabilities into code repositories, toolchains, and business systems, transforming "intelligence + capability" into a closed loop of executable tasks, rather than just superficial differences in model switching or open-source availability.
In terms of market mechanisms, the early popular general agents (including Xiaolongxia, Workbuddy, etc.) were essentially educating users, allowing those unfamiliar with prompts or not accustomed to using Codex or Claude Code to experience "AI helping you work" through a visual agent interface; once users are educated, who remains will depend on model capabilities and cost structures, rather than a single UI innovation.
Source: Public Information
ABAB AI Insight
A Few General Agents Dominate
Structurally, general agents are a category with strong scale and network effects:
The stronger the model + the more plugins, the more task scenarios can be covered, and users have no reason to maintain multiple "general brains" simultaneously.
Once leading agents establish a clear advantage in experience and ecosystem, small and medium general agents will be viewed by users as "transitional products," making it difficult to achieve long-term retention.
This is similar to the history of cloud vendors and mobile operating systems: initially, there is a flourishing variety, but ultimately only a few dominant players with "ecosystem + infrastructure" remain, leaving little space for small, comprehensive commodity categories, with exceptions only for "extremely narrow vertical + extremely deep capabilities."
Plugin Ecosystem Determines Openness
The key to openness and closure lies not in whether it is open-source or whether models can be freely switched, but in:
Whether it can serve as an "operating system" that accommodates various tools and services.
Whether it leaves a genuine capability entry point for third-party plugins, MCP services, and Skill markets.
From the user's perspective:
Products like Claude Code and Codex, even if they default to using "in-house models," have sufficient openness as long as they can connect to APIs, Git, CI/CD, and local tools.
What truly locks users in is "only being able to use their own models, their own plugins, their own hosting," with only shallow integration with third-party services.
The reason agents can operate is that the plugin ecosystem has made "model → action" a reality: they can run scripts, modify code, call databases, and pull data analysis, rather than just being able to "chat + write some demos."
Early Agents Were Educating Users
The main value of early agent products was to help ordinary users establish the understanding that "tasks can be delegated to AI as teammates."
Once this understanding is established, user loyalty is not strong; users will naturally migrate to whoever has the smarter model, more comprehensive ecosystem, or lower costs.
From historical cases:
When Claude Code first emerged, it sparked a wave of "code agentification," but once Codex launched a more mature GUI and deeper IDE integration, it quickly captured a portion of high-frequency users.
Workbuddy provided an "out-of-the-box" agent entry for those who did not want to deal with prompts or understand model details, educating users on "task delivery methods" rather than binding them to a specific brand.
In other words: early agents served as the "free education department of the industry," paving the way for models and ecosystems with real moats in the future.
Models Are the Moat, Agent Experiences Will Converge
UI, processes, task scheduling, queue management, and memory mechanisms can all be imitated and iterated, leading to similar "task lists + statuses + explanations" for users.
The real differentiators are: the intelligence of the model, reasoning depth, execution stability, and how many high-quality tasks can be run at the same cost.
Typical phenomena:
Cursor's agent experience is already very mature, supporting multiple models and tasks, but the hottest topics remain Codex and Claude Code, as they have optimized the combination of "capability + price": subscription structure, call quotas, latency, and stability, forming a cost-performance advantage.
Once users have experienced a smarter model (with fewer misunderstandings, less meaningless work, and better self-explanation), they are unwilling to return to an agent that "makes basic errors," even if the latter has a better UI.
The conclusion is straightforward:
The agent layer is a "shell" used to host models and plugins.
The moat lies in model and computing power costs, while the ecosystem is in plugins and data sources.
Real Paths for Individuals and Small Teams: Compete on Plugins, Not Agents
Creating a "general agent": there is basically no chance. You cannot win against models, computing power, or ecosystems.
Creating a "vertical agent": if it only packages a layer of UI and a few ready-made skills, it is essentially competing on marketing, with a very weak moat.
However: there are still many opportunities to create plugins based on agents:
A large number of traditional software and SaaS are not designed for agents/LLMs: interfaces, permissions, and data structures are unfriendly.
There is still an adaptation layer empty in the chain from "agent → skill/MCP → traditional services": doing security, abstraction, multi-tenancy, multi-role, and multi-environment.
There are several feasible directions:
"Old World → Agent" Bridge Plugins
Encapsulate the complex APIs of certain traditional systems (ERP, MES, trading systems, risk control platforms) into a few high-level actions, allowing agents to call them safely.
The differentiating factors are: permission models, auditing, error rollback, and business logic; this is the high-value part of the skill layer.
High-Value, Multi-Step Workflow Plugins
Skills themselves are entry points, but the real product lies in the subsequent services:
For example, a "due diligence skill": the entry is just a command, while the backend consists of your scraping, parsing, and evaluation services.
The agent is only responsible for scheduling, while you manage the entire business pipeline.
Unified Abstraction Layer Across Systems
Provide agents with an "abstract world":
Such as a unified model of "projects," "tasks," "risks," and "assets," mapped to different systems.
These types of plugins can become "soft infrastructure" shared by multiple agents and models.
The threshold is not in "writing a skill," but in:
What services, data, and assurance mechanisms are actually connected behind the skill.
Whether it forms a business chain and cognitive chain that is not easily replicable by others.
ABAB News · Cognitive Law
General agents are ultimately just shells for models and ecosystems, with the moat in computing power and intelligence, not in task list interfaces.
Early popular agent products are tools for educating users, not the ultimate carriers; once cognition is established, loyalty is very low.
The real opportunities in the agent era lie in plugins and services: whoever can transform old systems into AI-callable capabilities will master the new structure.