Kun Jing's Genspark Column Updated Again: Multi-Model and Multi-Agent Are the Future
"The Idea of a Single Super Agent That Does Everything Just Sounds Good"
The fifth installment of Kun Jing's "Seeing AGI" column has arrived. This time, Jing zeroes in on a specialized question in AI: will a single model rule them all, or will a multi-model, multi-agent architecture ultimately prevail? After extensive hands-on work at the cutting edge of the technology, he is convinced the latter represents the true future. It comes down to an old, unshakable principle: let specialists do what they do best. Jing argues that pursuing an all-capable single model, even if achievable, would inevitably lead to mediocrity. A multi-model, multi-agent architecture, by drawing on both open and closed capabilities, delivers the optimal user experience.
Below is Jing's original piece, compiled by BlueRun Ventures. Enjoy reading:
"The future of AI lies not in finding a perfect model — but in orchestrating the best combinations of specialized models and agents to create experiences no single system could ever achieve."
As a strategic partner to leading AI companies like OpenAI and Anthropic, we have a unique vantage point. Our collaborations give me hands-on access to the most advanced models from each lab. Through building products that serve millions of users, and extensive pre-production testing, I've reached an unpopular conclusion: the "one model to rule them all" theory only holds up on paper.
The reality: combining the unique strengths of different specialized models with customized agents is what creates exceptional user experiences.
I know this challenges conventional wisdom. But before you dismiss it, let me walk you through the logic.
Let's start with what every AI practitioner already knows: no single model dominates every domain today. Through deep collaboration with top model teams, we understand the strengths and weaknesses of each AI system: OpenAI excels at deep research and creative writing, Anthropic's Claude shines in agentic reasoning and hardcore coding, Gemini leads in multimodal understanding (image/video/complex visual analysis), Grok Heavy specializes in large-scale complex reasoning, while the Kimi+Groq combination stands out for extreme speed and cost-effectiveness... I've found this phenomenon extends far beyond programming. Every model is a battle-hardened domain expert: OpenAI's hundreds of billions of parameters forge creativity and research depth, Anthropic relentlessly pursues safety and reasoning capabilities, Google Gemini restructures its very DNA for multimodal understanding — they represent different philosophical paths to intelligence. And product practice has taught me: forcing a single model to do everything is like asking an F1 race car to double as a cargo truck, SUV, and minivan — possible, perhaps, but doomed to mediocrity. The real magic happens when you combine specializations: in our agent hybrid system, creative writing goes to OpenAI, complex reasoning to Claude, visual analysis to Gemini — this isn't merely performance optimization, but orchestrating domain-optimal systems to give users a true taste of "superintelligence." When technical peers witness this architecture, the exclamation is always the same: "My god, this IS AGI!" — because the essence of AGI is not a perfect monolith, but the intelligent orchestration of diverse expertise.
Imagine: you're trying to solve a complex problem on your computer using only basic tools — a text editor, calculator, and web browser. Contrast that with having a fully equipped workstation — an integrated development environment (IDE), databases, analytics tools, design software, and specialized applications tailored to each task. The gap in efficiency and output quality between these two scenarios is vast. Yet many in AI cling to the belief that "one model + a few general-purpose tools" is sufficient — simply because AI can search, write code, and build tools. This thinking puts the cart before the horse. Real-world professional domains operate on ecosystems of highly diverse, specialized tools. In a surgical suite, hundreds of specialized instruments are arranged — each meticulously crafted for specific procedural steps. In a film studio, professional lighting, sound, camera, editing, and post-production effects equipment line the walls. On a trading floor, real-time data feeds, risk management systems, quantitative trading platforms, and compliance monitoring tools flow in concert. It's the depth and diversity of these toolsets that produce professional-grade results — something a handful of general-purpose tools could never achieve. Look at how top professionals actually work: they invest enormous energy in building and maintaining the finest pre-made tools in their field. When a new task arrives, they immediately reach for the optimal tool to solve it. Even when they have the ability to "build a solution from scratch," they only invest resources in creating new tools when the emerging need is significant enough. This model maximizes efficiency, quality, and flexibility — it's the unshakable principle that leads to excellence. The AI industry will eventually recognize this pattern. The most successful AI applications won't try to conquer the world with a handful of basic tools — they're building complete toolsets that match the complexity and specialization of their target domains. Companies that understand this will deliver superior AI experiences — but note, not "general-purpose" AI experiences.
The theory of "one super agent that handles everything" sounds appealing, but it misunderstands the path to expertise. You wouldn't hire an MIT PhD to simultaneously serve as chef, driver, developer, and designer — even if they possess the basic capability to perform these tasks, it's neither economical nor optimal.
The economics are straightforward: specialization drives efficiency and quality. Imagine an MIT PhD making hamburgers — they could produce a decent burger, but at what cost? You're paying PhD-level wages for work that a regular cook could do faster, better, and at a fraction of the cost, while wasting talent that could be applied to specialized domains.
This isn't just about cost — it's about unlocking exponential capability through intelligent specialization. Top companies don't hire generalists to do everything; they assemble teams of specialists who excel in their respective fields. A world-class restaurant succeeds not because one person can do it all, but because the head chef, sous chef, pastry chef, and sommelier each contribute their expertise.
The real breakthrough isn't in a single agent's capabilities — it's in the orchestration layer that enables agents to communicate seamlessly, hand off tasks, and build on each other's contributions, collectively achieving goals no single agent could reach. Like a championship team, the magic lies not in individual stars but in the essence of teamwork: anticipating movements, covering weaknesses, amplifying strengths, creating possibilities that only collaboration can unlock.
Let's address the elephant in the room: some assert that model companies will eventually close their APIs, reserving their strongest capabilities for themselves — making multi-model collaboration impossible. This concern is understandable, but it overlooks a crucial economic effect:
Markets always reward whoever delivers the best user experience, regardless of a company's own business needs. The fundamental question is: which path can truly deliver exceptional experiences to users? I suspect you already know the answer.
The reality about the AI ecosystem: it's vast and diverse enough to accommodate multiple paths in parallel. In any thriving ecosystem, some will choose to build closed systems, while others remain open. This coexistence isn't a flaw — it's the hallmark of a vibrant ecosystem. Just as the tech industry has both Apple's closed ecosystem and Google's open ecosystem, serving different user needs and thriving in parallel, the AI world will evolve a similar balance. Some users prefer the seamless integration of closed systems; others favor the flexibility and power of multi-model architectures.
Crucially, this diversity of paths creates competitive pressure that drives innovation. Companies pursuing vertical integration will continuously push the boundaries of single-model capability; companies pursuing collaboration will constantly explore what collective intelligence can achieve. Both paths will produce breakthrough results that ultimately benefit users.
Multi-model, multi-agent systems embrace this reality and turn it into competitive advantage. By intelligently integrating top-tier capabilities from all sources — whether open APIs, partners, or proprietary models — such systems can consistently deliver experiences that surpass any single model. Users get the best creative writing, the best logical reasoning, the best visual understanding, the best coding assistance — all in a seamless experience.
The path I'm betting on: start from user needs and work backward to the optimal technical solution; rather than starting from internal capabilities and expecting users to accommodate limitations. In a world where open and closed systems will coexist, the true winners will be those who navigate this ecosystem intelligently, delivering ultimate experiences to users — whether that experience requires one model, or the symphony of thousands.
Genspark CEO Kun Jing: Don't Want to Be Replaced by AI? Here's What I Recommend You Start Doing Now
Genspark CEO Kun Jing: I See AGI, and I Hope I'm Wrong | BlueRun Family Headlines
Stop Asking Me What an AI Agent Is | BlueRun x Volcano Engine Event Recap
