What This Is

DeepSeek recently launched Harness — an Agent (an AI capable of autonomously executing multi-step tasks) runtime framework. Its core design is "everything is a plugin": models, tools, capabilities, even runtime modes are all essentially pluggable modules.

A developer ran a demonstration: feeding both DeepSeek's official plugin development documentation and the arXiv API documentation to the model, and asking it to write an arXiv search plugin. The result: the model read the documentation on its own, broke down the tasks, wrote the code, ran smoke tests, and ultimately delivered a working plugin — even generating usage instructions along the way.

What deserves attention is the second step. When an AI framework itself is fully plugin-based, and AI can also write its own plugins, the cost structure of tool development gets rewritten. Previously, building a customized AI tool required people who understood architecture, APIs, and deployment; now, clearly describing the requirement is enough for the model to deliver.

Industry View

Supporters argue this effectively amounts to DeepSeek turning the Agent framework into "Lego." SMBs can assemble AI workflows tailored to their own business on top of Harness, without reinventing the wheel from scratch. The model's ability to write its own plugins also lowers the cold-start threshold for the plugin ecosystem.

Dissent is plentiful. First, the demonstrator temporarily couldn't access Harness's GitHub documentation page, suggesting the infrastructure may not yet be ready to handle traffic. Second, "AI writing AI plugins" remains a toy-level demo — stability, error handling, and security boundaries in complex business scenarios are all unverified. Third, whether the plugin ecosystem can take off ultimately depends on whether there are enough real business scenarios — being able to write plugins does not equal demand for them.

Some developers caution: the demo's effectiveness is highly dependent on the underlying model's code capabilities. Swap in a weaker model and the results could be completely different. What gets flattened is the threshold for top-tier models, not for all models.

Impact on Regular People

  • For enterprise IT: when Agent frameworks go fully plugin-based, customizing an AI assistant tailored to your own workflows may shift from a "project" cost down to a "configuration" cost.
  • For individual professionals: in the future, everyone may have their own AI plugin library — automatically tracking new papers in a given academic field, auto-organizing customer emails, and so on.
  • For the consumer market: niche, vertical AI tools may see an explosion, because the development threshold has been flattened by AI itself.