DeepSeek Harness has connected to QQ bots through Tencent’s official QQ plugin. In testing, it can organize information, generate PPTs and write website code, but its design and interaction capabilities are not significantly better than comparable products. We believe the competition among personal agents is shifting from “who is smarter” to “who can users control and continuously modify.”
What This Is
“Big Fat Fish” is not a new model, but a type of electronic pet application built on DeepSeek Harness. Harness, an agent runtime framework, connects models, tools and task execution, while Aurabot adds personality, memory, mood and voice features on top.
This integration uses Tencent’s official QQ plugin and connects directly to DeepSeek Harness through the QQ Bot WebSocket API, eliminating intermediaries such as ordinary QQ accounts and communication bridge programs. Users can issue tasks in natural language through QQ, allowing the agent to search the web, organize information and deliver PPTs or code on its own.
Its open-source and plugin-based architecture is more significant. Model providers, tools, long-term memory and interfaces can all be replaced through plugins. A company can, for example, connect its own systems through MCP, the open protocol that enables models to access external tools and data sources. However, “everything is a plugin” does not mean every component can be replaced without cost: interface dependencies and version compatibility remain real burdens.
Industry View
We believe this release sends a clear signal: personal agents are evolving from fixed applications into composable infrastructure. Enterprises can retain internal data and existing systems while replacing models or adding tools. Developers can study and modify the execution process instead of having to accept a black-box product.
Counterarguments deserve equal attention. Open source lowers the barrier to experimentation, but it does not directly solve problems involving stability, permissions, security or maintenance. Once an agent can send messages, browse websites, generate files and even operate external systems, the risks of incorrect execution, prompt injection and sensitive information leaks expand along with its permissions. Moreover, the PPTs generated in our test were only moderately attractive, and the website designs were fairly conventional, demonstrating that there remains a gap between “completing the task” and “delivering a usable result.”
Impact on Regular People
For enterprise IT: Agents may become a new software entry point, but identity authorization, data boundaries and plugin compatibility must be addressed first. Being able to invoke an agent does not mean it can be safely connected to production systems.
For workplace productivity: News aggregation, first drafts and simple code can already be delegated to an agent in QQ, but the results still require human review, particularly for figures, citations and externally published content.
For the consumer market: The appeal of electronic pets will continue to lower the barrier to using personal agents. For now, however, its value lies more in a novel experience; whether it can become a long-term reliable “digital assistant” will depend on memory accuracy and task success rates.