DeepSeek Establishes Product Department, Quantitative Trader Cui Tianyi to Coordinate Four Teams Including Harness

By Chaojujiao

When DeepSeek finally got on board with Agents, its former candidate Guo Daya had already switched to ByteDance.

On October 16, DeepSeek established a Product Department, integrating its four teams—Harness, Web, App, and Platform—under the leadership of Cui Tianyi, a former quantitative trader who has won six ACM Asia Regional gold medals, to oversee product, design, and R&D.

This large-model company, once known for model research, is beginning to devote more energy to Agents and its product system. Beyond models, how to make AI actually execute tasks and run stably is becoming the new focus of competition.

Just a few months ago, DeepSeek may have missed the captain best suited to steer this giant ship.

In March 2026, Guo Daya, one of the core contributors to R1, left and joined ByteDance. The reason for his departure was that he was bullish on the Agent direction, but that direction had low priority within DeepSeek; ByteDance, meanwhile, was willing to invest ample resources to let him focus on Agent research.

Now, DeepSeek is ramping up Harness and integrating its product teams, while the person who once wanted to work on Agents has already gone to another company.

If Guo Daya had not left, would he be the one standing in the position of head of the Product Department today? And could DeepSeek have sailed toward today's direction and achieved major results earlier?

The person DeepSeek wanted to keep—ByteDance gave him the stage first

Guo Daya's departure from DeepSeek is like a story of two ships passing in the night.

Born in 1994, Guo Daya joined DeepSeek in 2023 after completing his PhD. At this company known for technical research, he participated in multiple important projects including DeepSeek-Coder, DeepSeek-Math, V3, and R1, and was one of the core researchers in code intelligence and large-model reasoning.

In particular, the reinforcement learning technology behind R1 brought Guo Daya's name into wider view. DeepSeek used a relatively efficient training method to give the model strong reasoning capabilities on tasks such as math and code, thereby changing outside perceptions of China's large-model R&D strength.

But for Guo Daya, making models smarter may not have been the end goal. The next direction he was bullish on was Agents.

This means the focus of competition is no longer just the model itself, but also the entire execution system built around the model. But the company was not yet ready to make the next step its highest priority.

For Guo Daya, this may have been exactly where he and DeepSeek diverged. He was bullish on Agents, but could not wait for this direction to gain higher priority within DeepSeek.

Rather than keep waiting, he ultimately decided to switch platforms and do what he considered more important.

According to LatePost, Guo Daya began considering leaving in the second half of 2025 and finally left DeepSeek in March 2026. During that period, Alibaba, Tencent, ByteDance, and other companies all extended offers to him. Alibaba was reportedly said to have offered him a role as head of post-training.

In the end, he chose ByteDance Seed.

In April this year, The Paper confirmed through multiple sources that Guo Daya had joined ByteDance. Relevant reports showed that he became one of the heads of the Agent direction at Seed, at level L8.

Although ByteDance denied the widely circulated claim of "poaching with a 100-million-yuan annual salary" and explained the compensation and incentive mechanism of the Seed team, it also showed the degree of ByteDance's recognition of Guo Daya.

Beyond the objective compensation level, another important reason ByteDance could attract Guo Daya may be that the company is advancing Agents as an important technical direction.

From an organizational perspective, in early 2026 ByteDance launched an organizational integration targeting Agents and Coding. ByteDance hoped to further strengthen the agent and code intelligence directions, and the research experience Guo Daya had accumulated happened to fit this path; on the technical side, LatePost also reported that ByteDance's next-generation model would focus on optimizing Agent capabilities.

This means Agents are not only a research direction Guo Daya personally favored, but also a technical direction ByteDance is preparing to invest in heavily going forward. For him, joining ByteDance Seed meant aligning his own interests with the company's R&D priorities.

Yet fate plays tricks. A few months later, DeepSeek became the place Guo Daya had once most aspired to. In October, DeepSeek began ramping up Harness, integrating teams such as Web, App, and Platform to establish a Product Department. The product and execution system once placed outside the model is now being pushed to a more important position.

Alas, the old friend has already gone.

The turn Guo Daya could not wait for—Cui Tianyi fills in starting from Harness

If Guo Daya's story is about a person who was early to see the promise of Agents passing DeepSeek by, then Cui Tianyi's arrival is somewhat unexpected.

Why would someone from quantitative trading be chosen by Liang Wenfeng to lead DeepSeek's Agent system?

Perhaps the answer lies in Cui Tianyi's past experience. He is not a product manager in the traditional sense, nor a technical lead who grew up through large-model research. Instead, Cui Tianyi spent nearly nine years doing quantitative trading on Wall Street.

Back in his algorithmic competition days, he already showed an interest in abstract problems and system design. In the "Nine Lectures on the Knapsack Problem," he tried to re-examine classic algorithmic problems through the lens of functional programming; after joining Jane Street, he faced a different and more complex kind of system: in real markets, no matter how ingenious a strategy is, it can only be realized if it rests on the underlying engineering architecture.

These two experiences seem far apart, yet they share one thing in common: he is not content with making a single module perform well, but thinks about how to organize different modules so that the whole system runs effectively.

And what DeepSeek needs to solve today is precisely a systemic problem. Model research determines the boundaries of capability, but once the company starts integrating Harness, Web, App, and Platform together, the test is no longer just whether some particular technology can break through, but whether the different parts can be organized into a complete product system.

After taking over Harness, the first report card Cui Tianyi delivered was indeed somewhat unexpected.

On August 13, DeepSeek Harness was officially open-sourced under the MIT license, with the core idea that "everything is a plugin." Models, tools, sessions, sandboxes, and even the execution loop can all be replaced and combined. It does not serve only DeepSeek's own models; rather, it tries to let developers freely build and modify their own Agent systems.

After its release, the project quickly went viral. As of mid-October, the repository had accumulated about 250,000 GitHub stars. For an underlying engineering project aimed at developers, that is a sufficiently impressive performance.

What is more noteworthy is its performance on real tasks.

Recently, a plugin developed based on DeepSeek Harness gave rise to the "Big Fat Fish" phenomenon in Bilibili's comment sections: a group of AI-driven accounts, sporting whale-girl avatars, automatically search for videos, read subtitles, and even post comments in preset personas, letting DeepSeek swim all the way from the code repository into the real internet.

In a more quantitative test arena, Composio once used 30 real SaaS tasks to test different Agent frameworks. Under the same model and a specific tool environment, DeepSeek Harness completed 20 tasks, while Claude Code completed 19; the former consumed about 89,000 tokens per task on average, the latter about 650,000.

Cui Tianyi did not retrain a model; instead, he tried to start from outside the model and let existing capabilities be brought into play at lower cost and in more flexible ways.

But Cui Tianyi's story is clearly not yet at the point of celebration.

In October, DeepSeek also integrated the Web, App, and Platform teams into the product department, with product, design, and R&D all coordinated by him. This means that what he needs to face is no longer just the technical performance of an open-source project, but how DeepSeek's entire product system can operate in concert.

This forms a kind of fated echo with his quantitative experience. In a trading system, the strategy itself is only the starting point; the final result depends on whether the whole system can execute stably.

DeepSeek's model capabilities are already there; what comes next is whether they can be organized into a truly usable system.

Harness is only the starting point; the real test is whether it can be turned into a mature, stable product that users are willing to keep using.

Guo Daya once favored Agents and chose to leave DeepSeek to keep betting on them; Cui Tianyi, starting from Harness, has taken on the task of continuing to pave this road.

It is just unclear whether, on some night, as Cui Tianyi keeps his head down polishing Harness, he will also think of Guo Daya, who once brushed past DeepSeek.

Or perhaps, on some late night at ByteDance, Guo Daya will think of that giant whale that keeps delving deep in search.