DeepSeek Harness: 1.54 Million Users, But Experimental Version

DeepSeek claims that its AI agent environment, Harness, has 1.54 million daily active users and 35,000 third-party plugins. These figures, which are unaudited and focused on the official API, come with a warning from the publisher: the software is experimental, lacks security audits, and is not presented as production-ready. A marketplace is planned to structure the ecosystem, although no date has been announced.

DeepSeek presents Harness as experimental, without audit

DeepSeek describes Harness as an experimental version intended for developers, without a security audit, and recommends not to consider it production-ready. The publisher specifies that protective mechanisms and approval requests may reduce risks but do not eliminate them. Risks arise from the permissions granted to agents and their extensions: a malicious plugin, an erroneous command, or hidden instructions in a document can, depending on access, lead to data leaks or file modifications.

Claimed records but unaudited and limited to the API

In an announcement on November 8, DeepSeek stated that it had reached 1.54 million daily active users as of November 4, 2026, less than three months after its launch on August 13 and the release of a desktop application on September 29. The company also mentioned a daily consumption of approximately 92 trillion tokens. This total does not equate to that many human messages or tasks completed. DeepSeek provides these figures itself, which have not been subject to independent audit, and notes that the user count pertains to those accessing its official API, without necessarily covering all deployments with other model providers.

An agent architecture where "Everything is a Plugin"

Harness structures the activity of agents that can, subject to permissions, access files, use tools, and execute commands to chain various operations. The stated principle is "Everything is a Plugin": models, tools, interface elements, and components are interchangeable or combinable. Extensions such as dsh-media-desk allow for the management of scripts, visuals, and storyboards, dsh-context is used for analyzing the agent's context, and agentero-dsh is utilized for processing scientific articles. These plugins are presented by their creators and by DeepSeek as examples, without comparative validation in a laboratory setting.

35,000 plugins, a marketplace announced, and MIT-licensed code

DeepSeek claims 35,000 third-party plugins and indicates that nearly 60% of active users have created or installed at least one plugin, which does not imply that they know how to develop them. The company plans a marketplace to facilitate the discovery and distribution of extensions, as well as a support program that includes token credits, although no specific timeline has been provided at this stage. The number of extensions alone does not inform about their quality or maintenance. Harness benefits from being published under the MIT license and can be configured with multiple models, contributing to the appeal of an open ecosystem. OpenAI is also developing agents capable of pursuing tasks, and Anthropic is creating specialized agents integrated into enterprise tools; modularity remains the main distinguishing feature of Harness, without any demonstrated superiority in usage. Additionally, a "super agent" project presented in early November aims to coordinate multiple autonomous assistants; this is complementary research and not proof of widespread functionality. DeepSeek aims to make Harness an operating system for AI agents, with the idea that tools installed by developers could remain in place even when changing models, a direction presented as potentially more sustainable than a benchmark record.