Intuit Rebuilds Its AI Agent Architecture Twice in Four Months

What happened?

Intuit, the company behind TurboTax, QuickBooks and Mint, has been a pioneer in the use of agentic AI. However, its path to success is far from linear. According to Nhung Ho, Intuit's vice president of AI, speaking at the VB Transform 2026 event, the company rebuilt its AI agent architecture twice in roughly four months.

Initially, Intuit implemented a system with a fleet of specialized agents, each designed for specific tasks. But customers complained about having to decide which agent to use for each task. The solution was a central orchestrator that automatically routed requests. This system worked for about three months, but then began to fail for a structural reason: agents passed results in natural language, and each transfer lost context, forcing the next agent to infer conclusions, degrading accuracy. With 10 agents chained together, errors accumulated inevitably. As Ho said: 'If you have 10 agents and they all pass information to each other, every time that pass happens, the error compounds.'

Faced with this diagnosis, Intuit abandoned the orchestrator and adopted a skills-and-tools-based architecture. Instead of independent agents, capabilities are now broken down into smaller, reusable components that can be shared across different parts of the product. The complete rebuild took 60 days, with a first working version in less than 20.

Why is it important?

This case illustrates the real challenges of scaling agentic AI in production. Many companies are adopting similar architectures, and Intuit's stumble offers valuable lessons. The orchestrator's failure shows that coordination between agents via natural language can be fragile and prone to compounded errors. Intuit's solution — moving to a skills-and-tools system — suggests that modularity and reusability are key to avoiding excessive complexity.

Moreover, the speed of the rebuild (60 days) demonstrates that it is possible to pivot when there is a clear vision and management support. Ho convinced leaders with a demo based on real customer queries, and engineers by arguing that shared skills would serve more users than isolated agents.

What consequences will it have?

For Intuit, the most visible result is a feature that allows a conversation with an AI agent to include a human at any time. Currently being tested with 1% of customers, this feature aims to improve the support experience. In the long term, the new architecture should enable smoother integration of AI into products like QuickBooks and TurboTax, automating complex tasks without losing accuracy.

For the industry, Intuit's case serves as a warning: there is no silver bullet in agent architecture. The choice between specialized agents, orchestrators, or skills systems will depend on context, but Intuit's experience suggests that simplicity and transparency in information transfers are critical.

What should readers know?

• Agentic AI is not a destination but an iterative process. Even pioneers like Intuit have to rebuild their systems.
• Communication between agents in natural language can be a hidden source of errors. It is better to standardize interfaces or use structured formats.
• Adopting a skills-and-tools architecture may require cultural changes: teams shift from building agents to running evaluations (evals) to measure performance.
• The key to organizational success is having concrete evidence (demos with real data) to convince both leaders and engineers.

Intuit's case reminds us that in agentic AI, iteration speed is more important than initial perfection. Rebuilding in 60 days can be faster than maintaining a flawed system.

In summary, Intuit's experience is a case study on navigating the complexity of multi-agent systems. The company not only survived two rebuilds but emerged with a more robust architecture and lessons that can benefit the entire industry.