Building a Personal Agent on DeepSeek Harness: Start with the Official QQ Plugin
Tencent's official plugin removes the need for third-party QQ bridges, so DeepSeek Harness becomes a drop-in personal agent base with a real chat front end. The larger consequence is ownership: because every component is a plugin, the memory, model, and tools that a viral bot ships bundled can be replaced piece by piece, turning a toy into a controllable agent.
The viral "Big Fat Fish" QQ pet runs on DeepSeek Harness, and that same open-source runtime is now the base for a DIY Personal Agent. Rather than adopting the ready-made Aurabot, the harness gets wired to QQ through Tencent's official dsh-qqbot plugin, which talks directly via QQ Bot WebSocket and skips the usual NapCat/OneBot bridge. Installing the plugin and scanning a QR code puts an agent in chat that can fetch and summarize AI news, generate a PPT, and write an introduction website from a single natural-language request; DSH execution logs show each turn calling the DeepSeek API.
Those tests don't push past what OpenClaw or Hermes did earlier this year. The appeal is architectural: DSH is MIT-licensed, and its "Everything is a Plugin" design lets model providers, tools, MCP integrations, memory, and UI all be swapped or extended. That positions the harness as a buildable base rather than a finished pet. The planned series covers digital twin, long-term memory, event awareness, goal planning, workflows, multi-agent coordination, multimodal interaction, and self-evolution.
A comparison with OpenClaw and Hermes shows the practical payoff of DSH is not raw intelligence but a pluggable architecture; the same model and tools can be owned and extended rather than locked inside a packaged pet.
Going with Tencent's official plugin instead of a community bridge points to agent frameworks getting first-party platform channels, which shortens the path from chat app to real tasks.
Without explicit UI or design direction, the generated website and PPT fall back to conventional output, so an agent's output quality still depends on how precisely the operator specifies the request.