Intuit Scrapped Its Own AI Agent Architecture Twice in Four Months; AI VP Calls It the 'Fast Path'

Intuit Scrapped Its Own AI Agent Architecture Twice in Four Months; AI VP Calls It the 'Fast Path'

In a revealing session at VB Transform 2026, Intuit’s VP of AI, Nhung Ho, detailed the company’s radical decision to rebuild its AI agent architecture twice within a four-month span. Rather than viewing these iterations as setbacks, Ho characterized the aggressive pivot as the "fast path" to achieving a reliable, production-ready system capable of handling complex financial tasks.

Why Multi-Agent Systems Broke Down

Intuit’s initial approach mirrored the industry-standard "multi-agent" trend: spinning up a variety of specialist agents—such as writers, planners, and reviewers—and connecting them via a central orchestration layer. However, the company discovered that this design suffered from a critical structural failure: context loss during natural language handoffs.

According to reports from VentureBeat, when agents passed results to one another using natural language, each downstream agent had to infer how the previous one reached its conclusion. This inference degraded with every "hop," causing errors to compound by design rather than by accident. As Ho noted, this effectively meant that a ten-agent chain was destined to fail.

The 60-Day Pivot to Skills-Based Architecture

Faced with these persistent failures, Intuit chose to scrap the orchestrator and decompose its monolithic agents into discrete, granular components. The company moved to a skills-and-tools-based architecture where a single planner coordinates base-level capabilities.

• Speed to Rebuild: The engineering team completed a first working version of the new architecture in under 20 days, with the full rebuild finalized in 60 days.

• The "Blast Radius" Fix: By separating the "brain" (the planner) from the "hands" (the specific skills), Intuit ensured that if a tax-filing skill misfired, the system could catch it and retry without compromising the payroll or finance modules.

• Engineering Buy-in: As highlighted by VentureBeat, Ho successfully convinced leadership and engineers by demonstrating the new architecture’s superior performance against real production customer queries.

Human-in-the-Loop and Future Scaling

The redesign intentionally brings human experts—such as accountants or product support staff—back into the workflow at critical checkpoints. This pattern serves both practical and regulatory needs, as seen in the company's early testing of a feature allowing human professionals to join live agent conversations with full context.

As reported by VentureBeat, the rebuild also transformed how Intuit gathers user feedback. By systematically analyzing the volume of feedback directly from chat interactions, the team can now identify system shortcomings at a scale that manual review processes could never match. This focus on modular, verifiable components represents a significant shift in how enterprises are thinking about the future of autonomous agent design.