DeepSeek's Most Promising Project Takes the Stage, Everything is a Plugin. Harness isn't trying to imitate anyone, but to give the power of definition to the open-source community.
On the evening of August 13th, DeepSeek finally broke its silence, releasing two important pieces of information. First, the official version of DeepSeek V4 Pro was released, and simultaneously updated on the APP, web version, and API. Users can use the brand new V4 Pro official model by selecting "Expert Mode" via the APP or web interface; the API model name remains unchanged.
Second, the " DeepSeek Harness Team," with its black whale avatar, made its first post, announcing the official launch of the Harness Developer Preview version, open-sourced under the MIT license.
As previously teased, the V4 Pro official version and the self-developed Harness made a simultaneous debut. Since we have already focused on introducing V4 Pro before, this article will focus on analyzing why DeepSeek Harness is noteworthy.
And after today, whether DeepSeek will transform from a model company into a harder-to-define company.
01
The Same Model, A Different Shell Makes It Seem Like a Different Person
To understand why Harness is important, one must first grasp a counterintuitive fact: in the Agent era, what determines whether an AI can get work done is not just the model.
On August 6th and 11th, the agent tool company Composio conducted two experimental tests attempting to prove the value of Harness. Researchers connected the same DeepSeek V4-Flash model to eight different Harnesses—the execution systems wrapped around the model—and had them complete thirty multi-step tasks respectively. These tasks weren't simple Q&A; they required the Agent to access real applications like Gmail, Google Calendar, GitHub, Slack, call tools, and change application states. Each task was allowed to run for up to fifteen minutes, and success required passing all checks.
The gap was significant. The best performer, Pi Agent, passed twenty tasks; the worst, OpenCode, passed only fourteen. Out of 240 total runs across eight Harnesses, only 129 succeeded. Among the thirty tasks, only six were completed by all Harnesses. In terms of cost, for similarly completing sixteen tasks, the estimated cost per successful task for Claude Code, Codex, and DeepAgents was approximately $0.195, $0.081, and $0.045 respectively. For the same model, the difference was more than fourfold.
This means that the model sets the upper limit of capability, while the Harness determines how much of that ceiling is ultimately realized and at what cost. Throughout a long-running task, the execution system must constantly make judgments: what to keep and what to discard in the context, when to call which tool, whether to retry or change course when a tool fails, whether to trust the model's completion signal or verify it again. If any step is mishandled, even if the model has the right idea, it could fail at the last moment when actually modifying files or submitting code.
It's worth noting that this is one of the very few projects that DeepSeek opened for early closed beta since its rise to fame.
Several developers received Harness beta access before the official release. Some testers discovered that having the same V4-Flash model complete the same game within DeepSeek Harness, Reasonix, and Codex produced drastically different outcomes. This shows that when the model is identical, the tools, prompts, context organization, and execution strategies provided by the execution system can significantly alter the final product.
This is precisely why DeepSeek values its self-developed Harness. Internally, DeepSeek has believed that Agents need to solve the problem of continuous learning, ultimately enabling AI to accelerate AI R&D. Looking down this path, Harness is not just an add-on programming tool, but a workbench for models to enter real development tasks.
Furthermore, a company that built its foundation on low cost and model capability would find itself in a precarious position if task delivery, failure feedback, and developer interfaces remained long-term dependencies on others' systems. Even with a price advantage, it wouldn't control how many tokens a task ultimately burns, how many retries are needed, or when a task is truly considered complete.
02
Everything is a Plugin, DeepSeek Isn't a Replacement for Anyone
What exactly is Harness? The most convenient description in the market is " DeepSeek version of Claude Code." But those who have used the beta version generally find this label inaccurate.
According to DeepSeek 's official documentation, the core design philosophy of Harness is "Everything is a plugin." It is built on the Cordis plugin system. Models, tools, skills, sessions, sandboxes, storage, loops, schedulers, UI—all Agent capabilities are composed of plugins. Developers can select, replace, or extend any capability through configuration without modifying the Harness source code.
This is on a different scale from the usual understanding of plugins. VS Code plugins add features to the editor; Codex plugins add tools to the Agent. Harness plugins can replace the Agent's brain, toolbox, rules, even its entire interface.
According to multiple sources involved in the beta test, developers went wild with plugin-making during the beta period. "People aren't properly testing the beta anymore; they're running off to write plugins." Hundreds of plugins emerged within just a few days—some completely overhauled the work interface, while others implemented "cross-session long-term memory with background self-evolution" using pure plugins, enabling the model to periodically review its work records and compress temporary experiences into lasting knowledge.
Some also commented that open-source community governance might be another challenge DeepSeek will face next.
Another crucial design aspect is traceability. Harness uses an append-only session log. Everything the model sees—system prompts, chain-of-thought, tool calls and results, sub-Agent scheduling, every context injection—is fully recorded. Context compression does not delete the original history; it only changes what the model subsequently sees through replacement events. Developers can trace by source in the Trajectory view, supporting recovery, forking, retrieval, and replay. The official documentation summarizes this principle as "If the model saw it, it's recorded."
This reflects a typical systems engineering mindset. According to reports, DeepSeek Harness team leader Cui Tianyi graduated from Zhejiang University's Computer Science Department, worked at the quantitative trading firm Jane Street for nine years, and joined DeepSeek in March 2026, with Harness officially starting internal development in May. The core moat of high-frequency trading systems has never been about smarter strategies, but stable execution in extremely complex environments, exception handling, comprehensive log traceability, and controllable risk—exactly the missing piece for moving AI Agents from demos to production. One beta developer described its performance in long tasks as "steady as an old dog": tasks are automatically saved when interrupted, ready to resume next time without starting over.
It's worth noting that Harness hasn't confined itself within DeepSeek models' walled garden. It natively supports connecting to nearly 40 major models including Kimi , OpenAI, Anthropic, and Google. Perhaps others tend to build a ready-made Agent, but DeepSeek Harness is more like a set of components, allowing developers to decide for themselves what the Agent should be like and how it should work.
Of course, the official team remains clear-headed. The announcement states that, as an early preview version, "there are still many details to be improved and refined, and core plugins and basic APIs will also undergo rapid iteration." In fact, one of the media outlets involved in the beta, ifanr, marveled that it was hard to imagine the famously fast DeepSeek taking over half an hour to run a programming task. The v0.1 version number also indicates that this black whale has just surfaced, far from swimming freely.
03
From Selling Tokens to Delivering Results, the Ambition of the Black Whale
The emergence of Harness coincides with a shift in the entire AI industry.
Over the past few years, competition among large model companies has focused on parameter scale and benchmark scores. But entering 2026, the core capabilities of leading models have rapidly converged, gaps in single-turn QA have narrowed, while price wars have intensified. The ceiling for the pure token-selling business model has become clearly visible. The industry is gradually realizing that the core value of the AI industry lies not in model output, but in real-world application—with the same token consumption, idle chat yields minimal value, while fixing a bug or completing a feature development can create several times the commercial value.
DeepSeek 's new release signifies a move from selling compute power to delivering results. In the old model, customers paid based on token usage, regardless of whether the model actually solved the problem. In the new model, the anchor point for payment shifts from how much compute was consumed to what task was accomplished.
This DeepSeek Harness release is also highly praised by developers for another reason. DeepSeek does not define the Agent itself. Instead, it leverages the diffusion power of the open-source community to promote a Harness-based approach. This is a grander, yet more difficult, path.
Furthermore, Harness faces a mature market already validated by Claude Code and Codex, with the latter backed by the model iterations and commercial resources of Anthropic and OpenAI, holding a first-mover advantage. After open-sourcing, questions remain unanswered: whether the plugin ecosystem can spontaneously flourish, whether developers are willing to entrust production environments to a v0.1 version, and whether the domestic market's payment habits can support the new logic of "pay-per-result." DeepSeek itself has stated that core plugins and APIs will continue to evolve rapidly, meaning early adopters may bear the cost of interface changes.
On August 13th, releases happened on both sides of the Pacific. Elon Musk took the stage for Grok 4.6, calling it "intelligent, fast, and highly cost-effective" ; DeepSeek said nothing, just kept pushing updates.
Between the clamor and the silence lies a divergence of two business philosophies.
We want to call this DeepSeek 's "most promising project in history," not because it's the strongest today. V4 Pro still has gaps with SOTA models, Harness is only a v0.1 developer preview, market reaction to recent price adjustments needs observation, and ecosystem building is a long journey. Potential is potential precisely because it is not yet realized.
But if we stretch the timeline, from R1 to V4, from models to Harness, DeepSeek has taken each step in the same direction: compressing cutting-edge capabilities into affordable costs, then connecting affordable capabilities to implementable systems.
The black whale has surfaced. As for how far it can swim—that lies in a future belonging to the open-source community.
This article is from the WeChat public account "Phoenix Network Technology," author: Dale, editor: Dong Yuqing
