Why I Killed Our AI Search Product With 5 Million Users

I’m Kay, Co-founder and CTO of Genspark AI. Here’s something counterintuitive: despite our AI search product reaching over five million users, we decided to sunset it. Why? Because we believe traditional AI search is already becoming obsolete. Let me share our story of building AI search over the past nine months.

The Limitations of First-Gen AI Search: Pre-defined Workflows Don’t Scale

When Perplexity launched in late 2022, it ignited excitement about AI’s potential to transform search. Yet that first wave followed a rigid workflow:

• Analyze the query and expand keywords
• Retrieve top web results
• Rerank/Summarize them into a final response

While adequate for basic questions, this framework crumbled with more complex needs, like comparing technical solutions, planning multi-factor purchases, or handling in-depth research. In essence, it was like trying to navigate a maze with only fixed turns.

At Genspark, we initially built our own AI search engine on the same foundation. Then we layered on incremental improvements:

• Specialized data sources (Scholar, Financial, Travel, Product, and more) to enrich information.
• Parallel search that automatically triggered deeper investigation for complex queries.
• Cross-check asynchronous agents to verify statements too complex for quick, on-the-fly handling, later expanded into data-search and deep-research agents.
• Mixture-of-agents approach to combat hallucinations, with multiple agents verifying each other’s outputs.

Although these innovations significantly improved quality and grew our user base, we realized we were still shackled by a legacy design: a fixed, predefined workflow. For truly adaptive, context-rich problem-solving, we had to break free entirely. That realization led to our Super Agent—a new paradigm that integrates all of our hard-won improvements without being bound by a static sequence of steps.

From Fixed to Fluid: Engineering Adaptability in Genspark Super Agent

Instead of forcing every query through a fixed workflow, the Genspark Super Agent adapts to the problem at hand. It plans each step, utilizes the best tools or sub-agents, observes the results, and adjusts its strategy in real-time—often surprising us with its creativity. If one approach fails, the Super Agent smoothly transitions to another, emulating human problem-solving on a large scale.

This flexibility extends to both breadth (which data sources or APIs it pulls from) and depth (how many iterative refinements it makes). For simpler tasks, it won’t waste time on unnecessary steps; for more complex ones, it can keep digging until it finds a satisfying answer. It can also tailor its output to each user’s needs—whether that’s a direct answer, a Sparkpage (article), a presentation, generated images, interactive HTML pages, or even phone calls.

Orchestrating Intelligence: The Triad of LLMs, Tools, and Curated Data

1. Dynamic Orchestration & Model Steering

Genspark Super Agent coordinates eight specialized LLMs through our Mixture-of-Agents framework. Each model is steered beforehand to excel at its assigned role, from rapid response to deep analysis. This hierarchy ensures stability while maximizing each model’s strengths.

2. Specialized Tools & Sub-Agents

Our library of pre-engineered sub-agents spans from presentation generators and Python code executors to call-making modules. Each is optimized for reliability and efficiency, enabling the Super Agent to handle tasks like creating charts or developing interactive pages—all without boxing users into rigid workflows.

3. Curated, Trustworthy Data

The Super Agent accesses carefully verified datasets, distilled from high-quality web sources, trusted partners, and expert-reviewed repositories. Offline verification agents continually audit and refine this data. By prioritizing accuracy over sheer volume, we minimize misinformation and ensure reliability in our outputs.

Here’s our latest performance:

Check out these real examples you can jump into and continue:

• AI “Call for Me” to Make Restaurant Reservation
• Create a Minute-Long South Park Episode About Recent News
• Find Top Fashion Influencers’ Contacts and Send Emails

Lessons Learned: Less Control, More Tools

In our transition from a rigid AI search engine to a fluid, adaptive Super Agent, we discovered an essential principle: less control, more tools. Overly structured workflows limit creativity and depth, whereas allowing multiple specialized agents to tackle different aspects of a problem—and granting them the freedom to choose and switch among various tools—unlocks far greater capabilities.

• Less Control: Prescriptive processes often curb exploration and make it harder to adapt to novel challenges. Embracing more open-ended strategies fosters innovation and resilience.
• More Tools: Equipping agents with specialized modules for data retrieval, analysis, presentation, and communication empowers them to craft end-to-end solutions on the fly. This not only supports advanced use cases but also keeps simpler tasks lightweight and efficient.

This fusion of adaptive planning, diverse tool support, and vetted data makes Genspark Super Agent far more flexible and dependable than any conventional AI search product. This advantage was so significant that we chose to retire our own thriving AI search solution to focus on the future.

Check out our product at https://www.genspark.ai and compare it with other AI Search products. See the differences yourself.