Guo Daya Missed the Boat, So Cui Tianyi Took the Helm

When DeepSeek decided to fully commit to Agent development, former candidate Guo Daya had already moved on to ByteDance.

On October 16, DeepSeek established a product department, consolidating four teams—Harness, Web, App, and Platform—under the leadership of Cui Tianyi. A six-time ACM Asia Regional gold medalist with a quantitative trading background, Cui was tasked with overseeing product development, design, and R&D.

This model-centric company, once renowned for its research prowess, has begun pivoting toward Agent and product systems. Beyond model innovation, enabling AI to execute tasks reliably and efficiently has emerged as a new competitive frontier.

Just a few months ago, DeepSeek may have missed its ideal “captain” for this journey.

In March 2026, Guo Daya, a key contributor to the R1 project, left DeepSeek for ByteDance. His departure stemmed from his belief in the Agent direction, which lacked internal priority at DeepSeek. ByteDance, however, was willing to allocate substantial resources, allowing him to focus on Agent research.

Now, DeepSeek is ramping up Harness and integrating product teams, while the researcher who once championed Agent has moved elsewhere.

Had Guo Daya stayed, would he now lead the product department? Could DeepSeek have accelerated its current trajectory and achieved greater success?

01 ByteDance Secured the Talent DeepSeek Couldn’t Retain

Guo Daya’s exit from DeepSeek feels like a missed opportunity.

Born in 1994, Guo joined DeepSeek after earning his Ph.D. in 2023. At this research-driven company, he contributed to pivotal projects like DeepSeek-Coder, DeepSeek-Math, V3, and R1, becoming a core figure in code intelligence and large model reasoning.

His work on R1’s reinforcement learning technology brought him into the spotlight. DeepSeek’s efficient training methods enabled the model to excel in mathematical and coding tasks, reshaping perceptions of China’s large model R&D capabilities.

Yet for Guo, enhancing model intelligence wasn’t the ultimate goal. His next focus was Agent technology.

A model that answers questions is merely one facet of AI. Future AI products must decompose tasks, invoke tools, execute operations, and adapt based on results. This shifts the competition from models alone to the entire ecosystem supporting them. However, DeepSeek wasn’t yet prioritizing this evolution.

For Guo, this divergence marked a turning point. He advocated for Agent but saw it sidelined internally.

Rather than wait, he decided to pursue his vision elsewhere.

According to LatePost, Guo considered leaving in late 2025 and departed in March 2026. During this period, Alibaba, Tencent, and ByteDance extended offers. Alibaba reportedly offered him a leadership role in post-training development.

Ultimately, he chose ByteDance.

In April, The Paper confirmed through multiple sources that Guo had joined ByteDance, leading Agent development at Seed with a job level of L8.

While ByteDance dismissed rumors of a “hundred-million-yuan salary,” it acknowledged offering competitive compensation and incentives, underscoring its high regard for Guo.

Beyond salary, ByteDance’s commitment to Agent as a strategic priority likely attracted Guo. Organizationally, ByteDance initiated Agent- and Coding-focused integration in early 2026, aligning perfectly with Guo’s expertise. Technically, LatePost reported that ByteDance’s next-gen model would prioritize Agent optimization.

For Guo, joining ByteDance meant aligning his research interests with the company’s R&D focus.

Ironically, months later, DeepSeek transformed into the place Guo had once aspired to lead. In October, DeepSeek ramped up Harness, integrating Web, App, and Platform teams to establish a product department. Product and execution systems, once secondary to models, now took center stage.

Alas, the former colleague had already moved on.

02 While Guo Daya Awaited His Turn, Cui Tianyi Stepped In with Harness

If Guo’s story is about missing an early opportunity, Cui Tianyi’s rise was unexpected.

Why did Liang Wenfeng entrust DeepSeek’s Agent system to a quantitative trader?

The answer lies in Cui’s background. He is neither a traditional product manager nor a model research veteran. Instead, Cui spent nearly nine years in quantitative trading on Wall Street.

Even during his algorithm competition days, he gravitated toward abstract problems and system design. In Nine Lectures on the Knapsack Problem, he re-examined classic algorithms through functional programming. At Jane Street, he confronted even greater complexity: in real markets, even sophisticated strategies require robust engineering to succeed.

These experiences share a common thread: the ability to organize diverse modules into an effective system.

This is precisely the challenge DeepSeek now faces. While model research defines capability boundaries, integrating Harness, Web, App, and Platform demands not just technical prowess but the ability to unify components into a cohesive product.

Cui’s first achievement with Harness was remarkable.

On August 13, DeepSeek Harness was open-sourced under the MIT license, emphasizing “everything is a plugin.” Models, tools, sessions, sandboxes, and execution loops could be freely replaced and combined. Designed for developers, it aimed to democratize Agent system customization.

The project quickly gained traction. By mid-October, its GitHub repository amassed 250,000 Stars—a rare feat for infrastructure projects.

Its real-world performance was equally notable.

Recently, Harness-based plugins sparked the “Big Fat Fish” phenomenon on Bilibili: AI-driven accounts with whale avatars automatically searched videos, read subtitles, and posted comments in preset personas, bridging DeepSeek from code repositories to the live internet.

In quantitative testing, Composio evaluated Agent frameworks across 30 real SaaS tasks. Using the same model and tools, DeepSeek Harness completed 20 tasks, while Claude Code finished 19. The former averaged 89,000 Tokens per task, compared to 650,000 for the latter.

Instead of retraining models, Cui focused on cost-effective, flexible utilization of existing capabilities.

Yet Cui’s journey is far from over.

In October, DeepSeek merged Web, App, and Platform teams into the product department, with Cui overseeing product, design, and R&D. Now, he faces not just technical challenges but the task of coordinating DeepSeek’s entire product ecosystem.

This mirrors his quantitative trading experience. In trading, strategy is just the starting point; success hinges on the system’s stable execution.

DeepSeek’s model capabilities were already proven. The next hurdle was organizing them into a reliable, user-friendly system.

Harness was only the beginning. The true test was whether it could evolve into a mature product users would depend on.

Guo Daya had championed Agent and left DeepSeek to pursue it. Cui Tianyi, starting with Harness, took up the mantle.

One wonders: on late nights refining Harness, does Cui think of Guo, the one who missed his chance? Or does Guo, at ByteDance, recall the company he left behind?

- THE END -