Insider: DeepSeek Is Forming a Harness Team to Benchmark Against Claude Code

链捕手Опубликовано 2026-05-22Обновлено 2026-05-22

Введение

DeepSeek is reportedly forming a dedicated "Harness" team to develop a code agent product, directly targeting Anthropic's Claude Code. According to internal sources and a social media post by DeepSeek senior researcher Chen Deli, the team will focus on building "DeepSeek Code Harness." The initiative involves recruiting for key roles like Harness Product Manager and Harness R&D Engineer in Beijing. DeepSeek defines its approach with the core formula: Model + Harness = Agent. This signifies a strategic shift from merely offering a powerful coding model to creating the essential middleware that connects the model to real-world developer workflows. The Harness will handle context management, tool calls, task planning, file operations, code editing, terminal execution, and feedback loops. The move highlights that competition in AI-assisted coding is evolving from pure model capability to ownership of the developer workflow entry point. While DeepSeek has strong foundational models (e.g., DeepSeek-Coder series), it has lacked an integrated, productized agent experience. The popularity of a community-built project, DeepSeek-TUI, demonstrated developer demand for a Claude Code-like tool using DeepSeek's models, but also revealed the limitations of unofficial solutions. By building its official Code Harness, DeepSeek aims to leverage its unique advantages: direct collaboration with its model training team, control over APIs and design, the ability to create a data feedback loop fo...

Author | Wang Bo, Jiazi Guangnian

According to information obtained by "Jiazi Guangnian" from a source close to DeepSeek, DeepSeek is internally organizing a new Harness team focused on a code agent product, internally benchmarking against Anthropic's Claude Code.

Senior DeepSeek researcher Chen Deli recently confirmed this on social media, stating, "DeepSeek is organizing a new Harness team to work on Harness-related products and research," and explicitly said, "In simple terms, it's to benchmark against Claude Code and create DeepSeek Code Harness."

This is not an ordinary recruitment.

Recruitment information shows DeepSeek is opening two key positions: Harness Product Manager and Harness R&D Engineer, currently limited to Beijing. DeepSeek's Beijing office is located in the R&F Centre in Haidian District, close to Peking University and Tsinghua University. Officially, it's located in the "Centennial Beijing-Zhangjiakou AI Innovation Belt," while colloquially, it's also in the recently popular "Wang Huiwen Area."

Core Definition: Model + Harness = Agent

In the job description, a core formula is placed in the most prominent position:

Model + Harness = Agent

This statement can almost be seen as DeepSeek's internal definition for the next phase of productization: the model itself is merely the foundation for an Agent. Everything outside the model—context management, tool calling, task planning, file reading/writing, code modification, terminal execution, feedback collection, and evaluation loops—is the critical part that enables the Agent to truly integrate into workflows.

The job description further states: "We are transforming DeepSeek's cutting-edge model capabilities into leading Agent products. All work beyond the model itself falls under the scope of Harness." Additionally, the role will participate in the entire process of developing "DeepSeek desktop Agent products" and "define DeepSeek's understanding of Harness."

Jiazi Guangnian analyzes that DeepSeek does not simply want to create a code assistant plugin but is complementing the middle layer that connects the model to real-world workflows.

Over the past year, the industry has proven: strong coding ability does not mean developers will actually adopt it; a model's capability to write code does not mean it can consistently complete an engineering task.

What truly changes developers' workflow is not the Claude model alone, but Claude Code; not the GPT model alone, but Codex; not just a code response in a chatbox, but an engineering agent capable of entering the terminal, understanding projects, reading/writing files, running commands, fixing errors, managing Git, and calling tools.

DeepSeek's past strength was its model. Now, it is beginning to add that layer of "hands" on top of the model.

I. Why DeepSeek Emphasizes Harness

In traditional AI product contexts, "code assistant" usually refers to two types of products: one is a completion plugin in the IDE, and the other is code Q&A in a chatbox.

But the term repeatedly appearing in this DeepSeek recruitment is not Code Assistant, but Harness.

Harness originally referred to a "test harness" or "execution framework" in an engineering context. In the Agent context, it is closer to an external system that enables the model to truly act. The model is responsible for understanding, reasoning, and generation; Harness is responsible for connecting these capabilities to the real environment.

The job description mentions that this role needs to plan the DeepSeek Harness product roadmap, connect researchers, engineers, the open-source community, and end-users, and communicate deeply with researchers from the model training team to achieve co-evolution between the model and Harness.

This is a crucial point.

It indicates that DeepSeek's goal is not just to wrap a shell around existing models but to make the Agent product itself a part of the model's evolution. In the past, the common product logic for large model companies was: the research team trains a model first, and then the product team builds applications based on the model's capabilities. However, in the Agent era, this sequence is being disrupted. The product is no longer just an outlet for model capabilities but a training ground for those capabilities.

A code Agent failing on a real project might not be due to product interaction issues but to incorrect methods of long-context compression by the model; it might not be a problem with the tool-calling pipeline but instability in the model's task decomposition strategy; it might not be insufficient coding ability but a lack of continuous understanding of engineering constraints, test feedback, and user intent.

Therefore, the value of the Harness team is not just "building a product" but turning real development tasks into a feedback source for continuous model evolution.

II. Why Must DeepSeek Complement Code Harness?

DeepSeek placed early bets on coding capabilities. From DeepSeek-Coder to DeepSeek-Coder-V2, DeepSeek has continuously increased investment in code models, with improvements in supported languages, context length, and complex task capabilities. Its problem is not a lack of coding capability but that this capability has largely remained at the model layer and has not yet become a high-frequency product in developers' daily workflows.

The popularity of Claude Code proves one thing: The competition in AI Coding is shifting from model capability competition to competition for developer workflow entry points.

This is also a lesson DeepSeek must learn now. More subtly, before DeepSeek officially stepped in, the developer community had already created a version of "DeepSeek Claude Code" for it.

An open-source project called DeepSeek-TUI previously gained popularity in the developer community. It is a coding agent running in the terminal that can read/write files, execute Shell commands, search the web, manage Git, and coordinate sub-agents through a TUI interface.

The popularity of DeepSeek-TUI highlights two issues:

  1. Foundation Mindshare is Mature: The DeepSeek model already has a foundation in developers' minds for being a code Agent. Otherwise, the community wouldn't naturally develop Claude Code-like products around it.

  2. Official Layer is Missing: DeepSeek lacks not model attention but an official Harness.

For developers, the appeal of DeepSeek-TUI is straightforward: low cost, domestic availability, long context, and relatively low deployment barriers. Many domestic developers aren't unwilling to use Claude Code but are constrained by price, access stability, account systems, and enterprise compliance.

However, community projects also have inherent limitations:

  • No matter how active a third-party open-source project is, it's difficult to truly grasp the rhythm of internal model capability evolution;

  • It can adapt around APIs but cannot reverse-decide how the model is trained;

  • It can work on prompts, toolchains, and interaction optimization, but it's difficult to systematically inject massive real-task feedback into model improvements.

This is precisely the significance of an official Harness.

By creating its own Code Harness, DeepSeek possesses several advantages that community projects lack: collaboration with the model team, design authority over interfaces, closed-loop training data, access to internal real task scenarios, and the ability for long-term developer ecosystem operations.

The open-source community has already paved the way: developers indeed need a DeepSeek version of Claude Code. Now, DeepSeek is reclaiming this path to build its own core product.

And DeepSeek officially starting to recruit means it is finally preparing to step onto the field.

Chen Deli mentioned at the 2025 World Internet Conference Wuzhen Summit last November: "One of our company's core advantages is long-termism, adhering to the main line of frontier AI breakthroughs. In this process, we have also abandoned many side paths, not engaging in those short and quick side projects."

After the model war, the real Agent war has begun. This time, DeepSeek is complementing the most critical layer between the model and action—Harness.

DeepSeek is equipping its model with a pair of hands.

Связанные с этим вопросы

QWhat is the main objective of DeepSeek's new Harness team, according to the article?

AThe main objective is to develop a code agent product, specifically a code intelligent agent, internally benchmarked against Anthropic's Claude Code. The focus is on building the "harness" - the external system that enables the model to interact with real-world environments, tools, and workflows, moving beyond just the model's coding capabilities.

QWhat core formula is highlighted in DeepSeek's job description, and what does it signify?

AThe core formula is "Model + Harness = Agent." It signifies DeepSeek's internal definition for its productization path: the model is just the foundation of an agent. The crucial part that allows an agent to integrate into real workflows is the harness, which includes context management, tool calling, task planning, file I/O, code modification, terminal execution, feedback collection, and evaluation loops.

QWhy does the article suggest DeepSeek's move to build Code Harness is crucial, beyond just having strong coding models?

ABecause the competition in AI-assisted coding is shifting from pure model capability competition to a competition for becoming the entry point into developers' workflows. Products like Claude Code have shown that developers adopt tools that integrate deeply into their workflow (terminal, project understanding, Git, etc.), not just models that can generate code in a chat interface. DeepSeek needs to bridge this gap to turn its model strength into a high-frequency product.

QWhat is the significance of the community project 'DeepSeek-TUI' mentioned in the article?

ADeepSeek-TUI is a third-party, open-source terminal-based coding agent that gained popularity. Its significance is twofold: 1) It proves that developers already perceive DeepSeek's model as a solid foundation for a Claude Code-like agent, demonstrating mature developer mindshare. 2) It highlights a gap: the lack of an official, first-party harness from DeepSeek itself, which the company now aims to fill with its own team.

QWhat key advantage does an official DeepSeek Harness team have over community projects, according to the article?

AAn official team has several key advantages: direct collaboration with the model research team, control over API/interface design, the ability to create a closed-loop system for training data and feedback from real tasks, access to internal real-world scenarios, and the capacity for long-term developer ecosystem operations. This allows for co-evolution of the model and the harness, which community projects cannot achieve.

Похожее

Three Years Later: Looking Back at My Predictions About ChatGPT in 2023

Three Years Later: Revisiting My 2023 Predictions on ChatGPT In March 2023, shortly after ChatGPT's launch, I made 20 predictions about its future. Now, in mid-2026, I've used AI agents to fact-check each one against the latest data. Overall, most major directional forecasts were correct, with only one outright error (incorrectly stating GPT-4 had 100 trillion parameters). Key successes included predicting that RAG and retrieval architectures would become the standard for handling knowledge and hallucinations, that natural language interfaces (LUI) would create a massive new industry layer beyond the models themselves, and that China would develop viable large language models, significantly closing the performance gap with Western counterparts within about three years. Predictions about the absence of mass unemployment, the rise of a new "robot network" for agent communication, and ChatGPT not possessing consciousness also held true in their core arguments. However, the "devil was in the details." Errors frequently involved specific numbers, timelines, or overlooking distributional effects. I tended to overestimate the speed of adoption (e.g., for agent networks) while underestimating the ultimate scale of capabilities or costs (e.g., AI winning IMO gold without tools, or the extreme capital required for frontier models). Other misjudgments included: underestimating how AI would reinforce, not dissolve, information filter bubbles; incorrectly assuming AI-generated content would easily circumvent copyright (it has instead triggered record-breaking settlements); and misidentifying where value would be captured (it accrued overwhelmingly to the compute layer, like Nvidia, not just the application or model layers). Key lessons from reviewing these predictions are: 1) Directional and mechanistic insights are far more reliable than precise numbers or absolute statements. 2) There's a consistent bias to overestimate short-term speed but underestimate long-term magnitude. 3) Errors often lie in missing distributional impacts within a generally correct aggregate trend. 4) Predictions phrased with nuance and caveats aged the best. 5) Some fundamental debates (e.g., on machine consciousness or the ultimate value chain) remain unresolved even after three years. This exercise is less about scoring the past and more about establishing rules for clearer thinking about the next three years of AI.

marsbit4 ч. назад

Three Years Later: Looking Back at My Predictions About ChatGPT in 2023

marsbit4 ч. назад

Three Years Later: Looking Back on My 2023 Predictions for ChatGPT

Looking Back After Three Years: Revisiting My 2023 Predictions on ChatGPT In March 2023, shortly after ChatGPT's debut and before GPT-4's release, I made over twenty predictions about AI's future based on limited information and intuition. Now, in May 2026, I revisited those forecasts using an AI-driven analysis with 41 Opus 4.8 agents to cross-reference them with the latest data. The assessment used symbols: ✅ Correct, 🟢 Mostly Correct, 🟡 Partially Correct, ❌ Incorrect. Overall, the directional judgments held up well, with only one major factual error regarding GPT-4's rumored parameter size (incorrectly cited as 100T). However, nuances and degrees of accuracy revealed more. **What Was Largely Correct:** Predictions about mechanisms and directions proved accurate. The rise of RAG (Retrieval-Augmented Generation) as the standard architecture for combating AI hallucination was confirmed, as was the transformative potential of LUI (Language User Interface) in creating a new industry layer atop GUIs. The emergence of "robot networks" (agent-to-agent communication protocols) and China's rapid catch-up in developing capable large models (closing the performance gap with top models to ~2.7%) were also on point. The analysis affirmed that LLMs lack consciousness and that the Turing Test merely measures perceived intelligence. **What Was Off Target:** Errors often involved specific numbers, over-optimistic timelines, or misjudged distributions. The prediction that value would primarily accrue to the application layer was half-right but missed NVIDIA's dominance as the profitable infrastructure layer. Forecasts about AI circumventing copyright issues and fostering a "global common ground" by averaging human viewpoints were incorrect; instead, major copyright settlements occurred and AI personalization is increasing. Estimates for model training costs ("$5-10 billion cap") were significantly off, underestimating frontier costs and overestimating replication costs. The notion that LLMs could never do complex math without tools was disproven by later models winning IMO gold. **Key Patterns from the Review:** 1. **Direction over precision:** Judgments about mechanisms and trends were more reliable than specific numbers or definitive statements. 2. **Timing bias:** There was a tendency to overestimate short-term speed but underestimate long-term magnitude and transformation. 3. **The distribution blind spot:** Aggregate-level correctness often masked uneven impacts (e.g., on young professionals' employment). 4. **The value of qualifiers:** Predictions framed with caution (e.g., "reportedly," "for now," "prototype in 2-3 years") aged better. 5. **Some debates continue:** Issues like the nature of "emergent abilities" or machine consciousness remain unresolved. This three-year review highlights that while seeing the big picture is crucial, humility regarding specifics, timelines, and disparate impacts is essential for future forecasting.

链捕手6 ч. назад

Three Years Later: Looking Back on My 2023 Predictions for ChatGPT

链捕手6 ч. назад

AI Bubble Warning: AI Investments Are Negative Returns for Most Tech Giants

The article issues a stark warning about a potential AI investment bubble. It notes that while the AI boom shares similarities with the TMT bubble of the late 1990s, its scale is vastly larger, currently driving 93% of U.S. GDP growth. Major hyperscale cloud providers like Microsoft, Alphabet, Amazon, Meta, and Oracle are planning to invest trillions in AI data centers over the coming years. However, calculations based on analyst projections for 2025-2030 reveal a concerning math problem: expected capital expenditure growth far outpaces projected revenue growth. Even under an extremely optimistic scenario of zero costs, the implied return on investment for most of these tech giants (except Amazon) is deeply negative. This suggests that the current trajectory could lead to one of history's largest shareholder value destruction events. The piece outlines two potential escapes: AI generating vastly more revenue than currently anticipated—a near-impossible task—or a significant cutback in the planned investment splurge. The latter scenario could trigger a domino effect, severely impacting the entire tech supply chain (from Nvidia to TSMC), potentially pushing the U.S. economy into recession, and causing a major stock market downturn. The author suggests upcoming high-profile IPOs by companies like OpenAI and Anthropic might represent a transfer of risk from early investors to public market participants. While the peak of the hype cycle might sustain investment through 2026, the fundamental financial dilemma remains unresolved, setting the stage for a potential market correction in 2027 or 2028, similar to the years following Alan Greenspan's "irrational exuberance" warning.

marsbit7 ч. назад

AI Bubble Warning: AI Investments Are Negative Returns for Most Tech Giants

marsbit7 ч. назад

From Tokens to Machine Labor: AI is Shifting from Tool to "Worker"

The article "From Token to Machine Labor: AI is Evolving from Tool to 'Worker'" argues that the business model for AI is shifting beyond simply selling computational resources (tokens, GPU hours) or model access. Instead, a new "machine labor market" is emerging, where the core economic transaction is the purchase of economically useful work directly performed by software. The central thesis is that AI pricing will evolve through four stages: 1) raw tokens, 2) standardized LLM capabilities (e.g., text generation), 3) industry-specific labor markets (e.g., legal review, radiology), and finally 4) a programmable results market where tasks like resolving a support ticket are bid on and priced based on outcome. In this future, buyers will care less about *which* model or GPU completes a task and more about whether the work meets specified standards for accuracy, latency, and cost. This transition reframes the impact of AI on human labor. Rather than simple replacement, it suggests a re-coordination where machines handle standardized, verifiable work, freeing humans for roles involving oversight, context management, responsibility, and final judgment. In some cases, this "last 1%" of human input becomes more valuable as it enables the other 99% to be automated. Furthermore, as AI reduces the cost of work, demand may expand, creating larger markets (e.g., 24/7 customer service) rather than just cheaper versions of existing ones. The article concludes that while infrastructure (GPUs, models, tokens) remains crucial upstream, the market is converging on a simpler, tradeable unit: machine labor that can be defined, measured, priced, and procured based on contractible specifications.

marsbit7 ч. назад

From Tokens to Machine Labor: AI is Shifting from Tool to "Worker"

marsbit7 ч. назад

Xiaomi MiMo's 99% Price Cut is Not Marketing! Luo Fuli Posts on X to Refute Critics

The price of Xiaomi's MiMo-V2.5 series API has been permanently reduced by up to 99%, specifically for the "Input (Cache Hit)" cost, which covers users re-reading historical context in long conversations. MiMo's head, Luo Fuli, published a detailed technical blog to clarify that this drastic price cut stems from genuine engineering breakthroughs, not a marketing stunt or a simple price war. The core of the achievement lies in six key engineering optimizations. First, the model architecture adopts a Hybrid Sliding Window Attention (SWA), reducing the memory footprint (KVCache) to 1/7th of a traditional model. Second, a dual-pool memory management system actually utilizes these savings, allowing a single GPU to handle over 5 times more concurrent users. Third, an upgraded prefix caching mechanism achieves a cache hit rate of 93-95% for repeated reads, meaning most such requests bypass GPU computation entirely. Fourth, a self-developed distributed cache (GCache) utilizes idle SSD space on existing GPU servers, eliminating additional storage costs. Fifth, an intelligent scheduling system (LLM-Router) efficiently routes requests to maximize cache reuse and performance. Sixth, Multi-Token Prediction (MTP) accelerates the model's text generation ("output") side. Together, these systemic optimizations dramatically lower the real computational cost per request, enabling the 99% price reduction for cached inputs while reportedly maintaining positive gross margins. Luo Fuli's disclosure aims to shift the narrative from "price war" to a demonstration of substantive AI engineering progress.

marsbit9 ч. назад

Xiaomi MiMo's 99% Price Cut is Not Marketing! Luo Fuli Posts on X to Refute Critics

marsbit9 ч. назад

Торговля

Спот
Фьючерсы
活动图片

Популярные категорииright-arrow

Популярные тегиright-arrow