Liang Wenfeng needs his own Tibo: DeepSeek product department integration and the Agent path

Two months ago, Cui Tianyi was still asking developers to help find bugs in a new product; two months later, he took over DeepSeek's product department.

On August 13, DeepSeek Harness was open-sourced. This software lets models call tools, operate files, and complete continuous tasks. Cui Tianyi, then the team lead, indirectly announced progress on Weibo by soliciting feedback: "The current version is still very incomplete, please give us more suggestions."

On October 16, The Paper reporters learned from informed sources that DeepSeek would integrate Harness with the web, App, and open platform teams, with him in charge.

As of October 23, DeepSeek's official website recruitment page had listed "AI Product Manager" positions under the "Product Department." These positions are mainly divided into two directions: first, the AI product direction is responsible for the web, App, and open platform, requiring lowering the usage threshold and improving retention and conversion; second, the Agent Harness direction requires planning the product roadmap, continuously tracking users' real tasks, and collaborating with model training researchers.

DeepSeek's ecosystem also continues to grow. For developers deploying models themselves, DeepSeek is also trying to provide more tools.

On October 20, DeepSeek open-sourced the inference framework DeepInfer, opening up some of the engineering capabilities from its own services to help developers run models efficiently. On one hand, letting developers put models to use; on the other, using its own products to help users finish their work. These two types of product tasks that DeepSeek undertakes are expanding simultaneously.

For large-model companies, products, results, and users are becoming increasingly important.

Three months ago, OpenAI applied the technology of its coding tool Codex to ChatGPT Work, letting it execute tasks across applications and files. Tibo Sottiaux, who participated in creating Codex, and his team once had a coding tool take on more people's work, thereby changing ChatGPT.

Another signal comes from tax filing.

In the "AI and Employment" series of interviews published by tech media Platformer on May 26 this year, Boris Cherny, head of Claude Code development, recalled to host Casey Newton: "Someone specifically installed a terminal just to file taxes in it." The team saw uses beyond coding and built Cowork for more work tasks.

AI competition is raising its requirements: models must be capable, and also usable so people can get work done. DeepSeek's integration and hiring this time also push the question beyond model parameters: who chooses which tasks to prioritize serving, lets users finish them, and then brings the problems back to R&D.

Cui Tianyi's Agent Formula

Cui Tianyi's Weibo bio is very simple, a formula: Agent = Model + Harness.

This combination intuitively shows Cui's understanding of Agent.

Five days after DeepSeek Harness was open-sourced, a user's request consisted of only four words: continue the task.

On August 18, user oaixnah asked in the GitHub discussion area of the Harness project: the task was unexpectedly interrupted, and there were still messages queued; could a newly issued continue instruction skip the queue and be inserted directly into the conversation? The user wanted the work to continue, but the system first needed to resolve when this sentence should be sent in.

On September 10, community account Gniy7Ga replied under the post, saying the issue had been fixed in a new version. The release notes for v0.1.5-rc.1 published by DeepSeek that day also listed features such as general file upload, previewing results in the sidebar, and opening files with local applications.

Cui Tianyi's direct involvement can also be seen in the version records.

On September 29, in the release notes published by Harness on the code hosting platform GitHub, a new feature was listed: the macOS and Windows desktop clients can install the command tools required for managing plugins through the menu, without separately installing a runtime environment and package management tools.

After this update, the contributor marked was "@tianyicui." Clicking into the account homepage shows the name Tianyi Cui, the organization DeepSeek, and the pinned project is exactly DeepSeek Harness.

On July 31, when Cui Tianyi introduced the official version of V4 Flash on Weibo, he mentioned that the benchmark test for letting models call tools to complete coding tasks used the then-unreleased DeepSeek Harness.

As for how this system works, DeepSeek's official website shows a demonstration. In the demonstration, the user asks to make a Pomodoro timer plugin, that is, a focus timing gadget installed into Harness. The process is: the system checks the development guide, writes files, installs and verifies, and finally delivers a clickable timer.

In this demonstration, the respective roles can be seen. The model is responsible for deciding what to do next, Harness provides execution tools, and then sends the operation results back for the model to continue deciding. An AI system that can call tools and continuously handle tasks in this way is usually called an Agent.

And Cui Tianyi's Agent formula refers to exactly this kind of combination.

According to The Paper's report, Cui Tianyi graduated from Zhejiang University with a degree in computer science, worked for nine years at the quantitative trading company Jane Street, joined DeepSeek in March this year, and his Harness team was also formed that month. About half a year later, the product made by this young team he led was still in preview and iteration, and his responsibilities had expanded to the existing web, App, and open platform.

These entry points reach different users. The web and App directly provide chat services, while the open platform provides models through APIs (interfaces for other software to call).

As early as the release of V4 on August 13 When the Pro official version was released, DeepSeek adjusted its interface so that Codex could connect; on September 10 it released V4.1 Flash and announced that Agent products such as WorkBuddy and OpenCode had completed integration. The company is launching its own execution environment while continuing to let its models enter other people's tools.

DeepSeek has also reached the next step

Bringing down Token prices was once an important step for DeepSeek in reaching users.

And in the formula "Agent = Model + Harness," DeepSeek's Model strength is indeed obvious, but what about Harness?

On October 17, DeepSeek released the V4.1 Flash update. In "DeepSeek V4.1 Flash New Version: Wish-Style Programming Made Real with One Click," published simultaneously on its official website, the company said the new version improved the ability to "build complete projects from scratch" in areas such as game development and 3D modeling, and shortened "task runtime and output length." The announcement also recommended that users upgrade Harness so that the execution environment adapts to the new model.

In the Harness version notes released the same day, another change pointed directly at the use case: the default role in the SDK (the software development kit for developers to call this system) was changed from a fixed programming Agent to a general-purpose AI Agent suitable for general tasks.

In fact, DeepSeek had long since brought its models into different usage environments.

In January 2025, it released an App, putting the model on phones; in August this year, it opened up Harness, letting the model enter a work environment where it can operate tools and files. Today, the official website offers macOS and Windows desktop versions, showcasing documents, spreadsheets, code, scheduled tasks, and result previews. In such products, a task must go through a series of operations such as reading materials, calling tools, and generating files.

However, the official website still labels Harness as a preview version.

In the October 17 update, it addressed two specific issues: first, after the application unexpectedly exits, interrupted tasks will clearly show as stopped, avoiding the impression that they were completed normally; second, replies to Office tasks were adjusted to avoid referencing preview images that have already been deleted. These changes relate to whether users can know how far the work has progressed and whether they can find the delivered results.

At present, DeepSeek has revealed some goals in its job postings. For example, positions for web, App, and open platforms are meant to improve retention and conversion; for example, the Harness direction is meant to track real tasks and plan the product roadmap.

From conversation to execution has become a trend, and product-level improvements have also entered the details.

In September, Kimi's desktop work assistant Kimi Work added features such as partial modification of selected documents and Office historical version management; starting October 6, Claude Cowork moved new tasks for Pro and Max subscribers to the cloud, allowing users to continue across desktop, web, and mobile, though accessing local files still requires the desktop client to remain open.

While expanding its product team, DeepSeek is also advancing a new round of financing. According to Bloomberg on October 22, citing people familiar with the matter, the company's current financing round is about 100 billion yuan, with Tencent and CATL among the main investors. Reuters followed up the same day, saying the deal is entering its final stage and financial details may still change. Bloomberg also mentioned that DeepSeek is increasing investment in AI data centers and open-source models.

A programming tool once changed ChatGPT

Codex was once a well-performing internal test product at OpenAI. But after its public release, it did not quickly become a tool that users kept using.

In May 2025, OpenAI launched cloud-based Codex, letting it read code, modify files, and run tests in a remote environment.

Tech media outlet Every later interviewed team members including Tibo and reviewed why this release was coolly received. At the time, inside OpenAI, a unified development environment automatically handled the preparation work, but external users had to move access credentials themselves and install the required software before Codex could start working.

The problems that internal trials failed to expose in advance were not limited to this.

After a task began, users also found it difficult to correct direction midway, and the model capabilities at the time were still insufficient. Every wrote in its report that Codex's release attracted attention outside OpenAI but was rarely used continuously.

From late spring to summer 2025, Tibo took over engineering management for Codex and then adjusted direction: letting the Agent run directly on the developer's own computer, using files and tools already on the machine.

At that time, the team already had a starting point: the command-line version of Codex released by OpenAI in April. On this basis, Tibo prioritized rewriting the underlying program that reads files and runs commands, separating it from the interface where users enter instructions. In this way, the same execution program could support different interfaces. The cost of this choice was that new features for the command-line version progressed slowly for a time.

Subsequently, this program supported the editor extension, bringing Codex into the software developers usually use to write code. New problems also emerged: when users give the AI several tasks at once, how do they view progress, check modifications, and take over when needed?

On February 2, 2026, OpenAI released the Codex macOS desktop application. In this version, tasks are placed in different conversations by project, multiple Agents can work in parallel, and users can check changes item by item.

According to OpenAI's official explanation, the model can already continuously handle more complex tasks, but the original terminal and code editor are not suitable for letting people direct and supervise multiple Agents at the same time.

On July 9, OpenAI disclosed that Codex is used by more than 5 million people each week. OpenAI also said that more than 1 million people already use it for work outside software development. On the same day, OpenAI built Codex technology into ChatGPT Work, and the Codex app also began merging into the new ChatGPT desktop client.

At this point, the execution capabilities formed in programming tools further extended to ChatGPT's general work tasks.

On the other side, Anthropic also encountered something unexpected.

In the aforementioned interview, Boris mentioned that some users installed a terminal specifically to file their taxes. Although Claude Code can accept natural language instructions, a tool entry point like the terminal is something even some engineers are unwilling to use.

Users had already brought non-programming tasks into Claude Code. At this point, what the team needed to do was provide an easier-to-use entry point, so that more people could use its capabilities without first having to learn this way of operating.

On January 12, Anthropic launched the research preview Cowork, putting Claude Code's execution capabilities into a desktop app, and targeting knowledge work beyond programming, able to directly handle local files.

According to Anthropic's official tutorial, users can authorize it to access a folder, then give organizing requirements, and have it classify and rename files in the downloads directory. In other words, the instructions that previously had to be completed in a command window were placed into an interface more familiar to ordinary users.

However, after creating the new entry point, new problems emerged again.

On September 16, Anthropic announced the merger of Cowork with chat, gradually opening a unified conversational interface to users. Anthropic explained that in the past users had to first decide whether to go to chat, Cowork, or Design for visual work, and work already started in one entry point was difficult to carry to another. After the merger, users can assign tasks, edit documents or slides in the same conversation, and continue using existing context and tools.

Google, meanwhile, changed the way users receive materials.

In September 2024, the research tool NotebookLM launched the "Audio Overview" feature. This feature can turn materials uploaded by users into a conversation between two AI hosts, available for users to download and listen to. A month later, it also allowed users to specify discussion focus and adjust the explanation according to the audience's level of expertise. Thus, the same material gained a use case away from the screen.

Putting the three sets of choices together, one can see what the product teams actually changed:

From OpenAI to Anthropic, and then to NotebookLM, models and products are merging into a kind of capability.

Long tasks, DeepSeek's "new account"

However, DeepSeek cannot directly copy the product experience of other companies.

On October 20, DeepSeek announced the open-sourcing of its internal inference framework DeepInfer. It is understood that DeepInfer integrates high-performance operators and a deeply optimized inference engine, supports distributed inference, speculative decoding, SWA Bounded Replay, structured output and other capabilities, and is compatible with NVIDIA and Huawei Ascend platforms.

This open-sourcing means DeepSeek is handing over to external developers part of its engineering capabilities in areas such as computing models on chips and scheduling request execution. The first batch opened is the core capability of offline inference; building a complete online service still requires filling in components such as cache management and load balancing.

According to DeepSeek, the model determines the boundary of capabilities, while how to serve large-scale users at a reasonable cost is the core problem that the inference framework needs to solve.

Open source and inclusiveness have always been important labels attached to DeepSeek by the outside world.

On January 20, 2025, when DeepSeek released R1, it gave a highly impactful price: in US dollars, $2.19 per million output tokens, while OpenAI o1 was $60 in the same period, about one twenty-seventh of the latter.

Now, low-price inclusive options are increasing, and DeepSeek's own prices have also undergone adjustments.

In RMB, V4 Flash's output price in April was 2 yuan per million tokens, raised in August, and after the release of V4.1 Flash in September, lowered to 4 yuan during off-peak hours and 8 yuan during peak hours. The fixed version of Qwen3.7-Plus (2026-05-26) in the Alibaba Cloud Beijing region, when a single input does not exceed 256K tokens, also has an official original output price of 8 yuan per million tokens for regular real-time calls.

A cheap single call does not mean a cheap piece of work. The increased overhead of long tasks is specifically reflected in the V4.1 technical report published in September.

The technical report mentions: the model needs to handle increasingly longer contexts, and the burdens of computation, caching and transmission increase accordingly. Caching stores the computation results generated when the model processes context; reusing them can reduce repeated computation, but also takes up storage space.

The report also states that at the same context length, compared with V4 Flash, V4.1 Flash's global KV cache usage drops to about one quarter, and persistent KV cache to about one eighth.

According to DeepSeek's inference deployment note released on October 20, at present, based on DeepInfer, DeepSeek V4.1 Flash has achieved high-throughput, low-latency large-scale inference services, with a single cluster's service throughput of 155 trillion Tokens/day, and overall revenue of $38 per card per hour (at valley price).

Low-price inclusiveness was once an important advantage of DeepSeek, but long tasks increase the burdens of computation, caching and transmission, and also bring new cost pressure.

"If needed, we have no obstacle to making applications either." In July 2024, Liang Wenfeng, interviewed by "Anyong", had already answered questions about applications this way.

At that time, what brought DeepSeek attention was V2 and its pricing. Liang Wenfeng attributed the price cut to the cost reduction brought by architectural exploration. V2 reduced the operating burden by compressing the context cache and reducing the parameters actually used in each computation, while developers, after obtaining the model, could deploy and modify it themselves.

Compared with other products, users have formed a distinct perception of DeepSeek. This is a spotlight, and also a shackle. Beyond requirements such as experience, results, and efficiency, DeepSeek's Agent naturally carries different expectations. This is its inertia, and also what DeepSeek's "Tibo" has to face.

References:

1. "DeepSeek Integrates and Establishes a 'Product Department', Cui Tianyi Serves as Head", The Paper

2. AI Product Manager recruitment information, DeepSeek official recruitment platform

3. Cui Tianyi's Weibo homepage and DeepSeek Harness open-source post, Sina Weibo

4. DeepSeek Harness product introduction and demonstration, DeepSeek official website

5. DeepSeek Harness version update records: [v0.1.5-rc.1], [v0.2.0-rc.2], [v0.2.1-rc.1], DeepSeek official GitHub repository

6. User discussions and replies on DeepSeek Harness regarding "continue task", GitHub Discussions

7. "DeepSeek V4.1 Flash New Version: Wish-Style Programming Comes True with One Click", DeepSeek official website

8. DeepSeek App release announcement, DeepSeek official website

9. "Kimi Work Release Log", Kimi official help center

10. "Use Claude Cowork on web, desktop, and mobile", Anthropic official help center

11. "DeepSeek Raises $15 Billion in Challenge to Leading US Labs", Bloomberg

12. "DeepSeek wrapping up $15 billion fundraising to fuel AI race, sources say", Reuters

13. "How Codex Happened", Every

14. "Introducing Codex", OpenAI official website

15. "Introducing OpenAI o3 and o4-mini", OpenAI official website

16. "Introducing the Codex app", OpenAI official website

17. "ChatGPT is now a partner for your most ambitious work", OpenAI official website

18. "Claude Code's creator on the end of the software engineer", Platformer

19. Claude product release records, Anthropic official help center

20. "Delegating your first task in Claude Cowork", Claude Academy

21. "Claude Cowork and chat are now one Claude", Anthropic official website

22. "NotebookLM now lets you listen to a conversation about your sources", Google official blog

23. "New in NotebookLM: Customizing your Audio Overviews and introducing NotebookLM Business", Google official blog

24. DeepInfer project description and "DeepSeek V4.1 Flash Inference Deployment Q&A", DeepSeek official GitHub repository

25. "DeepSeek-R1 Release", DeepSeek official website

26. o1 model description and pricing, OpenAI official API documentation

27. "DeepSeek-V4 Preview: Entering the Era of Affordable Million-Token Context", DeepSeek official website

28. "DeepSeek-V4-Pro Official Version Launched", DeepSeek official website

29. "DeepSeek V4.1 Flash: Stronger, Faster, More Affordable", DeepSeek official website

30. "Model Invocation Pricing", Alibaba Cloud Model Studio official documentation

31. "DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression", DeepSeek-AI, arXiv

32. "Unveiling DeepSeek: A More Ultimate Chinese Technological Idealism Story | 36Kr Exclusive", Anyong Waves

33. "DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model", DeepSeek-AI, arXiv

This article comes from the WeChat public account "Boxed Meal Finance" (ID: daxiongfan), author: Chenuo, published by 36Kr with authorization.