Harness, DeepSeek's AI agent tool, hits 1.54 million users per day

What Harness is, and why DeepSeek is betting on it

A "harness" is the software layer that turns a language model into an AI agent: instead of just responding to a prompt, the system gains the ability to execute code, call external tools, and act autonomously on behalf of whoever programmed the task. This is the infrastructure behind products like Codex, from OpenAI, and Claude Code, from Anthropic, today's benchmarks in the coding agent market.

DeepSeek, a Chinese start-up based in Hangzhou, launched its own harness framework on August 13, 2026, with a desktop app arriving on September 29. According to a report by the South China Morning Post, the company said this Sunday (November 8) that Harness has already reached 1.54 million daily users, less than three months after launch.

The numbers behind the growth

The data, released by DeepSeek itself, show the scale of token consumption by Harness via the official API:

• 1.54 million daily active users in less than three months of existence;
• about 92 trillion tokens processed per day, as of November 4, 2026;
• that is equivalent to approximately 1.1 billion tokens consumed per second.

These are self-reported figures from the company, with no independent audit cited in the report, but they give a sense of the appetite for agent tools built on open models, the same strategy that has already made DeepSeek known outside China.

An "operating system" for agents, with third-party plugins

What DeepSeek is selling isn't just speed of adoption, it's architecture. Cui Tianyi, a former Jane Street engineer who joined the Harness team in March 2026, described the project's ambition in a social media post this Sunday:

Our goal is for DeepSeek's Harness to act as an operating system for the agent era.

Cui Tianyi, a former Jane Street engineer who joined the Harness team in March 2026

Cui compared Harness to a smartphone operating system: it comes with standard tools, like an alarm clock and a calculator, but lets developers replace native functions with third-party alternatives, the same way a user swaps out their phone's camera app. In practice, this translates into concrete numbers: about 35,000 third-party plugins have already been created since launch, and 60% of active users have built or installed at least one.

Among the examples cited in the report, there's a plugin that combines reading, highlighting, and translation of academic papers, letting the agent answer questions about the content, and another that shows developers which information the agent is using and how it changes over the course of a conversation. DeepSeek also announced plans for a marketplace where developers publish plugins and users discover them, with a support program that includes token credits, technical discussions, and promotional help.

OpenAI and Anthropic already had first-mover advantage

Codex and Claude Code didn't come out of nowhere: OpenAI and Anthropic used the advantage of having developed their harnesses ahead of the competition to build coding products now considered market benchmarks. This is the ground DeepSeek is contesting, in a move the SCMP itself describes as a new battlefront between Chinese and American AI giants.

The difference in approach matters for anyone following the sector: while Codex and Claude Code are closed, verticalized products within each company's ecosystem, DeepSeek's Harness relies on a platform model open to third-party extensions, with a planned marketplace and explicit incentives for plugin builders. It's the same ecosystem logic that already worked for mobile operating systems, now applied to AI agents.

What changes for developers in Brazil

In short: Harness arrives as a third option for an agent framework, more open to customization than the proprietary alternatives that currently dominate Brazil's devtools market.

For those who already use Codex or Claude Code day to day, the appeal of Harness lies in being able to extend the agent with third-party plugins without depending on a single vendor's closed roadmap, something closer to what the open-source ecosystem of IDEs and extensions already offers. If the plugin marketplace promised by DeepSeek gets off the ground with the token credit incentives mentioned, that could lower the barrier to entry for small teams testing integrations specific to the Brazilian stack, without the cost of building from scratch.

It's also worth noting the backdrop: DeepSeek is already known for publishing open weights for its models, like the R1 family, a strategy that lowers the adoption cost for those running their own infrastructure. If Harness follows this same logic of openness at the agent layer, and not just the model layer, the cost-benefit equation for engineering teams that currently pay for Codex or Claude Code could shift in the coming months.

What's still unclear

The SCMP report doesn't confirm whether the Harness code itself is open, only that its architecture allows third-party plugins, which is different from being open source. DeepSeek also created its first product division in October 2026, led by Cui Tianyi, according to the Chinese outlet The Paper, a sign that the company is only just beginning to organize the project's monetization.

Independent quality comparisons between agents built on Harness and the results delivered by Codex and Claude Code on real development tasks are still missing, and usage numbers continue to be the ones DeepSeek itself releases, with no external verification cited so far.