Intuit scrapped its agentic AI architecture twice in four months, revealing the practical challenges of deploying agent systems at scale. The company pivoted from a fleet of specialist agents to a central orchestration layer, then abandoned that approach when the orchestrator itself became too complex to manage.
Intuit VP of AI Nhung Ho explained the sequence at VB Transform 2026. The company's first redesign consolidated multiple specialist agents under a single orchestrator. This centralized approach initially seemed cleaner but created new problems. The orchestrator accumulated complexity faster than engineers could manage it, eventually degrading system performance and reliability.
The second rebuild took 60 days total, with a working prototype in under 20. Intuit shifted to a skills and tools-based architecture, decentralizing control away from a single bottleneck. Rather than agents coordinating through a central hub, the new system gives agents direct access to a shared library of capabilities and tools they can invoke as needed.
This pattern reflects a broader lesson in AI infrastructure. Early agentic AI implementations often assume that central coordination solves complexity. In practice, adding a smart orchestrator creates a new single point of failure. When the orchestrator must track state across dozens of tasks, manage error recovery, and handle edge cases, it becomes harder to debug and modify than the original specialist-agent approach.
Ho's presentation reframes these failures as progress. "The fast path," she said, involved learning what doesn't work through rapid iteration rather than trying to predict the right architecture upfront. Two major rewrites in four months looks chaotic from the outside but represented faster convergence than lengthy planning cycles would have produced.
The practical takeaway is straightforward. Agentic AI systems benefit from decentralized design patterns where individual agents retain autonomy and coordinate through tools rather than subordinate to orchestration logic. This mirrors successful patterns in distributed systems engineering,