DeepSeek's Harness Reaches 1.54M Daily Users

DeepSeek announced on Sunday (November 8) in a post on X that Harness, its open-source agent, achieved approximately 1.54 million daily active users on November 4, less than three months after its launch. Bloomberg reported these figures on the same day. These figures were disclosed by the company itself, without independent verification, and the count only considers users connected to DeepSeek's official API.

Adoption Curve

The series published by the company illustrates the speed of growth. The daily average was 271,000 users in the week starting September 30, increasing to 429,000 in the week of October 7, 665,000 in the week of October 14, 1.13 million in the week of October 21, and 1.40 million in the week of October 28. From the first point to the last, the number of daily users multiplied by approximately 5.7 over five weeks.

The volume processed accompanies the user base. DeepSeek states that Harness consumes about 92 trillion tokens daily, equivalent to approximately 60 million per active user. In an earlier update on the evening of November 5, the team had reported 5.48 million monthly active users as of November 4, according to specialized media reports. The ratio of daily to monthly users, close to 28%, indicates recurring usage rather than just curiosity around the launch.

Harness was released on GitHub in mid-August under the MIT License, which allows commercial use at no cost. The repository is currently in preview for developers, at version 0.1, according to documentation cited by specialized publications.

Plugins Without an Official Store

The company claims that Harness already has around 35,000 plugins, with 60% of users having created at least one, varying in complexity. DeepSeek plans to launch an official plugin store and a support program for creators, although a release date has not yet been set.

For a CISO, this number describes a software supply chain without curation. Agent plugins gain access to repositories, terminals, and credentials, and the recent experience with other code agents shows that the risk is not theoretical. Apiiro researchers found 2,539 distinct secrets exposed in 1,994 public GitHub repositories from configuration files of Claude Code, from Anthropic. The issue, therefore, is not specific to DeepSeek; it is a characteristic of the agent model with broad permissions, increasing with the installed base.

Reading Across Markets

In the United States, the competition for the layer orchestrating code agents reached a milestone in the same week. Meta acquired Conductor on November 4, a startup whose application runs multiple code agents in parallel and claims to have over 100,000 users, according to Axios. The metrics are not comparable, as Conductor speaks of total users and DeepSeek of daily active users, but the direction is the same: value is migrating from the model to the tool that distributes work among multiple agents.

In China, Harness serves as a distribution channel for DeepSeek's API. Since the daily user metric only counts those using the official API, each new user also translates to paid token consumption, and the 92 trillion daily tokens represent, in practice, the potential revenue that the free agent generates for the model. It is a business model of giving away the code and charging for inference, applied at scale.

For Brazilian, European, and Indian companies evaluating code agents, the decision now has two separate layers. The MIT License allows running Harness with other models, including on proprietary infrastructure, reducing dependence on the Chinese supplier. However, using DeepSeek's official API places source code and engineering data under the jurisdiction of where the inference occurs, a matter that risk committees of banks and consulting firms treat differently from a simple tool choice.

The number that deserves attention is not the number of users, but the number of tokens per user. Sixty million tokens per day per active user suggest agents running for extended periods without a human reviewing each step. When this volume reaches corporate contracts, the inference cost will no longer be a small line in the IT budget, and control over what each agent can access will no longer be optional.