What This Is

DeepSeek Harness went open-source this week, racking up 147,000 GitHub stars in 4 days and 6,600+ community-contributed plugins — but DeepSeek itself states in the popup that the current version targets developers, not regular consumers.

It is an Agent framework (a tool that lets AI complete multi-step tasks on its own). Its core pitch, what the team calls "Everything is a Plugin," is that all functionality lives as plugins. What can you actually do with it? Hands-on testing by developers found that users can have an Agent search the plugin marketplace and install plugins on its own, and have an Agent write and publish a new plugin. In other words, where AI used to be a "pre-trained assistant," DeepSeek now wants it to be "someone who assembles their own tools." The team also shipped a trace feature that lays out every tool call the Agent makes on a timeline — similar to Chrome DevTools' Network panel. Normal users find it maddening; developers find it addictive.

Industry View

Two numbers worth tracking: token consumption is brutally competitive. A tested in-depth session cost only 2.6 yuan, and the cache hit rate (the share of calls that reuse previously computed results) hit 98%. The team has already signaled upcoming price increases. This means the operating cost for complex applications like Agents is being driven to new lows by Chinese players.

But here is the sobering side: current version is 0.1, with no multimodality (no image or video input) and no scheduled tasks. In a test asking the Agent to write a clipboard-history plugin, UI debugging alone took over an hour of back-and-forth, and because image input was not supported, we had to describe bugs in plain text. There is still a meaningful gap between "it can do things" and "it can do them well."

Another risk worth watching: once Agents learn to find and install plugins on their own, where is the boundary on "installing the wrong plugin" or "installing a malicious one"? We have not seen any clear review mechanism disclosed yet.

Impact on Regular People

  • For Enterprise IT: Agent self-evolution is an imaginative direction, but stability, audit, and controllability are all still early-stage. It is not suitable for critical business flows within 2026 — better to let R&D teams pilot it internally first.
  • For Individual Professionals: If you are not a developer or AI practitioner, you can watch from the sidelines for now — no rush to dive in. If you are, use it to practice writing plugins and submit PRs; the next month or two is the window for building your position.
  • For the Consumer Market: DeepSeek has driven token prices down to this level, layered with an open-source strategy — forcing other Chinese model companies to differentiate on "capability" or "toolchain." How much they raise prices next will be the watershed moment.