Cómo Genspark alcanzó 250 millones de dólares de ARR tras reconstruir su producto alrededor de un Super Agente impulsado por Claude

Genspark launched out of stealth in mid-2024 as an AI search product, then progressively expanded into parallel search and asynchronous deep research that could run for minutes and return a finished analysis. The product kept growing because each expansion unlocked harder problems — but the architecture underneath was a directed graph of predefined workflow nodes, and that design was hitting a ceiling. Edge cases broke the workflow. Simple questions ran through too many steps. Hard questions hit walls the graph didn't know how to route around.

Co-founder and CTO Kay Zhu had been watching the ReAct-style agentic loop — where a model decides each next step rather than following a fixed plan — since 2023. The idea was elegant: if the model is powerful enough, it should know when to call which tool and when to stop. He tested it against every new frontier model, roughly every three months, for two years. The failures were always the same: infinite loops, hitting the same error repeatedly, or failing to recognize when it had enough information to stop.

In early 2025, Zhu tried again with Claude Sonnet 3.7. 'To my surprise, it was very intelligent,' he said. 'It knew when to stop and knew which tools to call. When some tool returned an error message, it would know what alternative way to try.' That recovery behaviour — understanding that an error is a signal to try a different path, not a reason to repeat the same action — was the missing piece that no prior model had reliably exhibited.

Genspark rebuilt the product entirely around it. The Super Agent that launched in early 2025 is model-agnostic by design, with different frontier models handling different tasks. Claude plays two central roles. First, the agent loop itself: at each step in a task, Claude decides which of more than 150 available tools to call next, when to gather more information, when to backtrack, and when to summarise and return. For the most demanding planning problems, Genspark fans out to multiple frontier models in parallel and uses Claude to reconcile their outputs into one coherent plan. Second, code generation: behind every AI slide deck, spreadsheet, and document that users see, Claude generates the code that produces the final artefact.

The results were sharp. Genspark surpassed $250 million in annual recurring revenue after pivoting to the Super Agent. 100% of the company's internal code is now written by AI — a reflection of how deeply the model has been integrated into every layer of the operation, not just the product.