DeepSeek published new figures for its Harness agent platform on the evening of 5 November. Measured across users on the official API, daily active users reached 1.54 million as of 4 November, consuming roughly 92 trillion tokens per day.
The platform now carries about 35,000 third party plugins. Around 60 per cent of active users have built or installed at least one.
The operating system claim
DeepSeek’s stated ambition is to be the operating system of the agent era: supply stable base capabilities and common functions, and leave professional scenarios, vertical industries and personal requirements to developers building plugins.
The design principle is stated as everything is a plugin. That includes functions that ship with the product, which third party developers are free to replace with their own versions. It is an unusually aggressive commitment for a company that also operates the underlying model, because it hands control of the user experience to people it does not employ.
Version 0.1 shipped on 13 August. In less than three months the project has moved from an agent tool into an application platform, and the community has already concentrated on content creation, design, context management and research workflows.
What 35,000 plugins actually cost
Scale brings a problem the team has acknowledged. Does a larger plugin library make an agent stronger, or merely more expensive to run?
Plugin descriptions and tool rules occupy context. Every additional integration competes for tokens that the model needs in order to reason about the actual task, and it adds latency and cost per call. For a platform already consuming 92 trillion tokens a day, context overhead is a direct operating expense, not an abstract concern.
The product question underneath is the one every open platform eventually faces. Should a general user receive the freedom to assemble arbitrary combinations, or a curated product that works out of the box?
DeepSeek has chosen freedom, and the 60 per cent adoption figure suggests users agree with the choice for now. It also means the platform’s quality is increasingly determined by third parties, which is a difficult position from which to promise reliability.
Why this matters beyond DeepSeek
The competitive frame for Chinese AI has shifted from model benchmarks to platform lock-in. A developer who has built three plugins against Harness has made a small but real switching cost commitment, and that is worth more to the platform owner than a marginal benchmark gain.
The open question is monetisation. DeepSeek has said it will begin issuing tokens to plugin authors, which turns an open ecosystem into a subsidised one. Whether that subsidy buys durable platform share or simply funds experimentation that migrates elsewhere is the number to watch in the next two quarters.
Editor’s note: This is an adapted translation of the original 36Kr report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://36kr.com/p/4057774236652417.