DeepSeek’s open-source agent app Harness reached 1.54 million every day energetic customers as of 4 November, counting folks on the corporate’s official API, and people customers burned by about 92 trillion tokens a day, the group mentioned in a neighborhood replace.
The figures appeared in an 8 November X article from the DeepSeek Harness account. That’s the first time DeepSeek has printed Harness utilization numbers. Dividing the token whole by the person depend works out to roughly 60 million tokens per energetic person per day — way over a standard chat session, and a plain sign of how a lot compute long-running brokers absorb once they learn recordsdata, name instruments and loop with out somebody on the keyboard for each step.
From August launch to a plugin ecosystem
DeepSeek says Harness v0.1 went stay on 13 August and a desktop app for macOS and Home windows adopted on 29 September. A harness, on this context, is the software program layer that turns a language mannequin into an agent — software program that may act on a person’s behalf relatively than solely reply to prompts.
The corporate frames Harness as an “everything is a plugin” stack. Since launch it counts roughly 35,000 third-party plugins, and about 60% of energetic customers have constructed or put in a minimum of one. It additionally notes that neighborhood plugins nonetheless fluctuate in maturity. The 1.54 million determine covers customers on DeepSeek’s personal API solely, so folks operating Harness towards different mannequin suppliers should not within the depend.
Market and writer grants, nonetheless to come back
DeepSeek says it’ll harden the plugin API, tighten plugin safety and construct an official market so builders can publish plugins and customers can discover them. It is usually making ready a help programme for plugin authors that would come with token grants, technical assist and promotion. No launch date or detailed guidelines have been printed but; a signup type is accumulating curiosity for now.
The identical week, Alibaba’s Co-Design Bench confirmed that even the strongest coding agent absolutely cleared solely 26.4% of actual app-edit duties — a reminder that agent adoption and agent reliability should not the identical race. Individually, Microsoft has warned that agentic tooling also can compress assaults on leaked cloud keys into minutes.
Why it issues
DeepSeek’s personal numbers put a Chinese language open-source agent stack previous 1,000,000 every day customers in below three months, with token use that appears extra like always-on automation than chatbot chatter. That’s early proof — from the corporate, not an auditor — that the business’s subsequent battle is as a lot about agent infrastructure and plugins as it’s about mannequin scores.

