Intuit overhauled its AI agent architecture twice in roughly four months, scrapping a multi-agent orchestration model in favor of a granular skills-and-tools design built for reliability, speed, and scalability. The changes affected the core business platform powering TurboTax, QuickBooks, and other financial products serving millions of customers across the United States. Intuit detailed the effort publicly at VB Transform 2026, where engineers outlined both the technical decisions and organizational challenges behind the pivot.
Intuit rebuilt its AI agent architecture twice in four months, prioritizing reliability over complexity across millions of users.
The rebuild effort traced back approximately three years, when internal testing indicated existing infrastructure would hit hard limits within two to three years as AI model sizes continued to grow. That warning prompted an initial architectural shift, moving from a fleet of specialist agents toward a central orchestration layer. That approach failed. The orchestrator accumulated complexity over time, transforming from a reliability layer into a primary point of failure. Agents passed results between each other in natural language, causing context loss and compounding errors at each handoff. Large, general-purpose agents handling broad capabilities introduced brittleness wherever workflows were complex or tightly constrained.
With the orchestration model proving unstable, Intuit initiated a second full rebuild. An initial working version was delivered in under 20 days, and the complete overhaul finished in approximately 60 days. The new architecture decomposed large agents into specialized components, explicitly separating planning responsibilities from execution. Rather than assigning broad autonomous behaviors, the design gave agents narrowly scoped skills and APIs. Orchestrator and planner components were rebuilt around explicit workflows, with skills and tools embedded down to the base level of the platform. The approach favored simplicity, robustness, and explicit control over agent actions, deliberately moving away from emergent behavior produced by complex multi-agent coordination.
Scalability was a central concern throughout the redesign. The legacy monolithic structure had created bottlenecks as AI capabilities expanded, limiting the system’s ability to handle fast queries and highly complex tasks simultaneously. The rebuilt infrastructure addressed this through modular processing layers that allocate compute resources dynamically based on task complexity. The architecture was also structured so that underlying models and providers could be swapped out without requiring a full rebuild of surrounding systems, reducing the cost of future changes as the AI landscape continues to shift.
Intuit’s experience reflects broader instability observed across the industry in multi-agent orchestration at scale. The failure patterns encountered reinforced a view that agent architectures demand painful iteration before reaching production-grade reliability. By replacing emergent coordination with explicit, modular control, Intuit moved toward a system where failure modes are more predictable and recoverable. The two rebuilds, completed within a compressed timeframe across a platform serving millions of users, represent one of the more documented examples of an enterprise-scale AI infrastructure overhaul driven directly by observed production failures. Nhung Ho is scheduled to present the technology decisions behind the new system at VB Transform 2026, including the introduction of an abstraction layer designed to decouple orchestration from specific model providers.
