How a ‘vibe working’ approach at Genspark tripled ARR growth and supported a barrage of new products and features in just weeks

Traditionally, product releases can be cumbersome, for which multiple sign-offs, endless crafts, bureaucracies and friction points require.

Genspark has developed a much different approach.

The Lean team of the AI workspace company practices the work of AI-Native or ‘atmosphere’, if you want to move them to what they call ‘gene speed’. This enables them to release new products and functions in rapid succession (almost every week or so), which increases the annual recurring income (ARR). As the Company hasIt can “be the fastest growing startup ever in terms of Arr.”

“When people work in the AI-Native way, everyone is actually the manager,” Kaihua (Kay) told Zhu, co-founder and CTO, to Venturebeat. “They are equipped with a team of AI agents, who are a kind of their report, and they are able to deliver end-to-end on his own.”

Aggressive rollouts, heating competition

Gensparklaunched in June 2024 by Main funerWas initially focused on AI search assignment. But despite achieving an impressive 5 million users, the company turned from that first product to Supermanwhich, instead of following a static order of steps such as traditional search, chooses the best tools or sub-agents for the task, the results of the meters and adapting in real time.

Super Agent will be launched on 2 April and is powered by the Claude of Anthropic and can condense an afternoon of a white collar office in 5 minutes, Zhu claims. For example, it can call, download, fact check, produce podcasts, produce concept documents, conduct deep research and bring together spreadsheets and slides.

“We still see it as a kind of search, but it is more technically advanced,” says Zhu, who has more than 20 years of experience to work at Google and Baidu.

The company has added more and more functions in the past four months; Here is an overview of his rolls and milestones:

• April 11: $ 10 million ARA only reached 9 days after the launch of Super Agent
• April 22: Introduced AI diseases (with hundreds of templates)
• April 28: A personalized super agent rolled out with adaptive personalities
• May 2: Save $ 22 million ARA, exactly one month after the launch
• May 8: rolled out AI sheets that make complete spreadsheets in one click
• May 15: introduced a fully agent download agent and AI Drive that manages and stores files
• May 19: Save $ 36 million ARR
• May 22: rolled out ai who can call
• June 4: Introduced an AI secretary who manages Gmail, Calendars and Google Drive
• June 10: rolled out an AI -Browser and MCP store with extensive browsing options and a tool market.
• June 18: introduced AI documents for making documents and management
• June 25: Introduced design studio with “canva-like” possibilities for making visual content
• July 10: rolled out AI pods to make podcasts with simple prompts
• July 17: Advanced editing functions introduced for AI -Dias
• July 31: rolled out AI -Dia’s 2.0
• August 1: Introduced multi-agent orchestration that can produce up to 10 agents at the same time

Genspark also heats the AI agent space with friendly competition. After OpenAi had announced it Chatgpt -agent In mid -July, GenPark carried out a comparative analysis and is “very confident” in his ability to perform the rival too much. To drive home this point, the company launched a “1 million dollars Side-by-Side AI confrontation“Challenging users to hunt for cases where other platforms perform better than Genspark Super Agent.

In the first round, users were instructed to build a financial slide of 12 pages with the help of Genspack and Chatgpt agent; Users identified 429 cases in which the latter performed the first, each earned $ 100 for their efforts.

In round 2 (which ended on Monday, August 4), Genspark increased the Ante to $ 200 per victory and opened the competition for an AI tool as an opponent. Users were challenged to use exactly the same prompt to build slides at Genspark and their chosen AI tool and then upload them to Gemini for evaluation.

“I don’t try to start a drama here – just really enthusiastic about how far the entire AI agent ecosystem has come”, the company Posted on X. “It shows that we all push the boundaries in the right direction.”

Some user reactions:

How Genspark’s AI Native Team Vibes

The Secret of Genspark is the slim, AI-Native Team of 20 people and engineering philosophy of “Less control, more tools.” Zhu explained that more than 80% of his code was written by AI, which is not atmospheric coding in itself, “because the coding of the atmosphere indicates that you never look at the code.” GenPark has previously has a “very rigid” code supply process to guarantee the quality of their code basis.

“We only need a very small AI-Native team to work in a kind of superhero mode, such as The Avengers“Said Zhu, who said they will gradually add team members if needed.” The AI coding and AI workflow are so powerful, it is a magnifying glass. “

Today’s Enterprise teams must be reorganized “completely different,” he said. He has managed 1,000 members teams with different levels of management and seen how office politics can introduce friction.

The Genspark team, on the other hand, communicates in ‘a very transparent way’ and productivity is ‘super high’. “Everyone is working on a product that can send,” said Zhu. “I believe that will be the norm that looks forward, because AI helps more and more people to do their job better.”

He also emphasized the importance of immersing yourself in your own product. From designers themselves to the marketing team: “We are actually eating our own dog food. We are our own product consumer. This way we will continue to improve experience.”

Inside Genspark’s flagship Super Agent

Zhu noted that when PerTlexity was launched in December 2022, the excitement ignored the potential of AI to transform search. Nevertheless, the rigid workflows followed, where platforms had to be:

• Analyze questions and expand keywords;
• Pick up top -web results;
• REALING/Summarize for a final response.

This was sufficient for basic issues, but “crumbled” in more complex scenarios such as technical comparisons, in-depth research and multi-step and multi-factor purchases. “In essence it was as if you were trying to navigate a maze with only fixed turns,” said Zhu.

GenPark built its search engine on the same type of foundation, placed on incremental improvements, including specialized data sources, in parallel searching for deeper research into complex questions and cross control of Asynchronous agents to verify statements too complex for “fast, on-the-fly handling”. But they realized that they were still “coupled” by fixed, pre -defined workflows, Zhu reported.

Super Agent uses nine differently, different specialized Great language models (LLMS) in a mixture of agents (tired) system. Models break up tasks in steps, delegate based on specialty and strength and then verify each other. Super Agent is also equipped with more than 80 tools (from sub-agents who can generate Python code to those who can call autonomously) and compiled more than 10 data sets from the internet, partners and repositories.

Genpark gives tasks to Claude, OpenAi, Google Gemini, Deepseek., AI’s Grok 4 and others: “Then let everyone produce their output, and we have an aggregate model to look through the results and analyze which process is the most cost -effective,” Zhu explained. “In this way we improve accuracy, we reduce hallucinations.”

The company also refines its own border model. However, they are not overly aggressive in creating ultramodern systems such as Deepseek V3 or V4, Zhu emphasized. The goal is to have the model performed at a low level, but heavy work.

“We are not trying to shift the border of the Frontier model,” he said. “We try to lower the costs and latency, because many own models are too large, too slow and too expensive for many relatively simple tasks.”

Regarding the trend of the atmospheric coding, the goal of Genpark is to have everyone experimented, even for non-programmers where the concept can be a bit ‘too far’.

“Many people think:” Vibe coding, I have heard about it, it sounds cool, but I am not familiar with the integrated developer environment (IDE), I am not familiar with code, “said Zhu.” With the help of Genpark, people can actually atmosphere. ”