DeepSeek announced that its open agent framework Harness, released in version v0.1 on August 13, 2026, reached 1.54 million daily active users as of November 4, 2026, with users consuming approximately 92 trillion tokens per day through the official API. The ecosystem has accumulated around 35,000 third-party plugins, used by approximately 60% of active users. The company announced an official plugin marketplace and token grants for third-party developers, positioning Harness as the "operating system of the agent era" and competing with OpenAI and Anthropic not only on model pricing. All key figures are currently data from DeepSeek itself without independent verification, while the community has already noted plugin overload, looping, and token overconsumption in the framework.

image
image

What happened

According to company data, Harness's audience grew at an accelerating pace: 271,000 daily active users were recorded for the week ending September 30, 2026, 429,000 for the week ending October 7, and DeepSeek recorded 1.54 million on November 4. Approximately 60% of these active users utilize third-party plugins, with the number of plugins in the ecosystem reaching around 35,000. Total consumption via the official API amounted to approximately 92 trillion tokens per day: this is roughly 1.1 billion tokens per second and an average of about 60 million tokens per active user per day. Following the adoption metrics, DeepSeek announced the launch of an official plugin marketplace and a support program for third-party developers with token grants.

Context

Harness is not a separate language model, but an open agent layer on top of LLMs: a framework where the model does not simply respond to prompts, but autonomously performs tasks, planning steps and calling tools until a result is obtained. It is precisely this agent loop with numerous intermediate calls that explains why inference load is measured in trillions of tokens per day and why autonomous tasks require orders of magnitude more computation than chatting with a bot. The plane of competition is also shifting: instead of a pure price war, DeepSeek is building an open plugin ecosystem layer where third-party developers expand the agent's capabilities, and the company itself gains a distribution and user retention channel. For open agent frameworks, such a network effect is rare, so the question is not only about model pricing, but also whether the ecosystem layer can become the main weapon against the agent stacks of OpenAI and Anthropic.

Why this matters for the industry

For the industry, the published volumes are a public indicator of the real scale of agent load on inference: long autonomous tasks require orders of magnitude more computation than chatbots, and this changes the economics of inference and capacity planning practices for everyone launching agent workflows. Competition between OpenAI, Anthropic, and DeepSeek is shifting from model pricing to ecosystem richness: DeepSeek is betting on an open plugin layer with a marketplace and grants, while the framework itself becomes a ready-made distribution layer for agent products, where value is created by plugins and agent scenarios, not thin wrappers around LLMs. For startups and builders, this is a real but time-limited window: before the launch of the official marketplace, third-party plugins remain a direct path to a multi-million user base, and after the launch, the battle will unfold in top categories—file handling, browser, data, and RAG connectors. If the trajectory holds, the agent layer will emerge as a separate competitive niche between models and applications: DeepSeek will try to establish itself as the open "OS of the agent era," while OpenAI and Anthropic will likely respond with their own ecosystem programs and integrations.

Why this matters for users

Harness can be tried right now: the framework is open-source, so it is worth installing and independently measuring how many tokens, time, and money your tasks consume, rather than relying on others' figures. The agent works differently than a chatbot: it autonomously performs a task rather than simply responding to a prompt, and thousands of third-party plugins significantly expand its out-of-the-box behavior—from file and data tools to browser scenarios and RAG connectors. However, the community has already noted downsides: when connecting 15 or more plugins, "plugin overload" begins with looping and context overflow, the same tasks cost approximately 20 times more in tokens and take tens of times longer compared to OpenCode, and updates may lose user configuration. Therefore, the framework should only be released to production with guardrails: plugin whitelists (significantly smaller than 15), strict token and time limits per task, loop detection, configuration backup before each update, testing updates on RC versions, and independent quality measurements instead of blindly trusting releases.

What is currently unknown / limitations

All key figures—1.54 million DAU, approximately 92 trillion tokens per day, and an average of 60 million tokens per active user—are data from DeepSeek itself, disseminated by SCMP, 36Kr, and KuCoin without independent measurement: arithmetically they are self-consistent, but the methodology for counting users and tokens has not been disclosed. Adoption metrics prove nothing about quality: the volume and coverage of plugins are metrics of distribution and engagement, not proof of superiority over OpenAI and Anthropic's agent solutions, and none of the sources contain agent quality benchmarks. It has not been published what share of consumption goes to solved tasks versus repetitions, errors, and looping: the 20-fold overconsumption compared to OpenCode remains an unofficial community measurement and allows for the possibility that consumption growth reflects layer inefficiency, not just task complexity. Finally, the question of retention remains open—whether users will remain active for more than one week—as well as whether the plugin marketplace and grant program will launch as announced and under what plugin verification rules.

Sources

Author

Look at AI, editorial team