Summary of Key Points
Recently, the focus of competition among domestic AI agents has shifted from “who has the smarter model” to the battle for Agent Harnesses – which can be considered as the “butlers” of these agents. These harnesses are responsible for managing the agents’ capabilities (Skills), tools they use, operating environments, cost settlement, and security measures. The essence of this competition is: who will have the control over what abilities the agents can use, how they perform tasks, whose resources they consume, and who is in charge of data security.
Detailed Explanation
1. Why has the Harness become the new battlefield? – The shift from “brain” to “butler”
Previously, the competition was about which models (such as GPT, GLM, etc.) were the most advanced. However, having just a model is not enough; agents need to be able to write reports, search for information, open documents, and format content for publication; they also need to interact with corporate knowledge bases and use office software. These tasks are not inherent in the models themselves and require coordination – this is where the Harness comes in. The Harness acts like a butler, organizing the models, skills (the methods of performing tasks), and tools, deciding which tools to use, how much resources to allocate, to whom data should be sent, and whether user confirmation is needed. For example, if you ask an agent to write for a public account, the Harness will arrange: use a search tool to find competitor articles, then an analysis skill to extract key points, followed by a writing tool to generate the content, and finally, a publishing tool to upload it to the public account platform. Whoever controls the Harness controls the actual operation of the agent.
2. Skills and Tools: one is the “instruction manual,” the other is the “actual executor”
- Skills (task guides): These are like “operation manuals” for agents. For instance, a skill for managing a public account might tell the agent to “select a topic → find competitor content → rewrite the text → format it.” But without the corresponding tools, these skills are useless. For example, if a skill suggests finding competitor articles but there’s no tool to retrieve the content, the agent can only provide suggestions and cannot execute the task.
- Tools (actual executors): These are the tools that enable the agents to perform tasks. Examples include search tools that can scrape web pages, Lark interfaces for document creation, and AI tools for image generation. The barriers to entry for these tools are clear:
- Publicly available tools (e.g., web content extractors) are easily replicable and often free or low-cost in the long term.
- Proprietary enterprise tools (e.g., internal knowledge base interfaces) and high-performance tools (e.g., video generation) are more difficult to replicate and have become focal points of competition due to their value.
For example, WorkBuddy’s public account management skill can coordinate the entire process (search → analysis → rewriting → publishing), but it relies on Tencent’s public account platform tools to actually publish the content. Without these tools, the skill would be ineffective.
3. Free vs. paid: cost and barriers determine the market
Manufacturers have categorized tools into free and paid options, with clear distinctions:
- Free tools: Basic skills (generic templates), local tools (that read local files), and open-source interfaces (such as MCP). These are low-cost or easy to replicate and are used to attract users. For example, Qwen-MM-Plugins’ local image reading function is free because it uses the user’s own computer’s processing power.
- Paid tools:
- Cloud-based high-performance tools (for image/video generation, real-time search) require manufacturers to cover server costs.
- Proprietary enterprise tools (corporate knowledge bases, internal system interfaces) require access rights and data security measures.
- Auditable tools (for modifying corporate documents, handling batch tasks) require logging and accountability.
For instance, Huoshan’s agent package is priced based on the number of credits used, including models, search capabilities, and generation services. Bailian’s exclusive API keys can only be used for specific agent tasks and cannot be used elsewhere – these tools are being sold to users as part of the service.
4. The future of Harnesses: integrated into everyday tools with a focus on security and control
Harnesses will not be standalone apps but will be embedded in commonly used tools such as Lark, WeChat, and cloud services. Future competitive areas include:
- Model selection: The harness will automatically choose the most suitable model for a task (e.g., Codex for coding, Stable Diffusion for image generation), without the need for users to make decisions.
- Tool integration: WeChat official accounts, Lark documents, and corporate email systems will all become tools that agents can use. The provider with more stable interfaces and clearer permission controls will have an advantage.
- Unified billing: Fees will no longer be based on the number of tokens (model usage) but on the total cost of tasks (e.g., the number of credits needed to write a public account article).
- Security first: Enterprises are concerned about agents potentially altering files or leaking data, so features like minimal permission levels, operation logging, and reversible actions will become key competitive factors.
In the future, when using Lark to write a report, the harness will automatically use search tools to find information, AI to generate a draft, and sync it to the team’s knowledge base – you might not even be aware of its presence, but it will have completed most of the work for you.
Conclusion
The competition among domestic AI agents has shifted from who has the best models to who can effectively manage their tools, operating environments, and cost management systems. Those who can control these aspects will gain an advantage in the future intelligent assistant market. For end-users, this means more convenient and efficient assistance – provided they are willing to pay for the high-value tools needed to achieve that.