Intuit Rebuilds AI Agent Foundation Twice in 4 Months

Intuit, a major US financial software company, completely rebuilt the foundational design of its AI agents (AI programs that autonomously operate for specific purposes) twice within approximately 4 months. Nhung Ho, Vice President of AI at the company, revealed the details at the 2026 'VB Transform 2026' conference. The final reconstruction took 60 days to complete, though the first working version was finished within 20 days.

Originally, Intuit employed a configuration with multiple 'specialized agents' tailored to different purposes. However, this approach required users themselves to determine which agent to use for which task, presenting usability challenges. The company therefore added an 'orchestration layer' (an intermediary management system that acts as a conductor) to automatically route tasks to appropriate agents. According to Ho, this configuration functioned for approximately 3 months.

However, the orchestration layer collapsed due to its complexity. When multiple agents passed processing results to the next agent, they exchanged information using natural language (everyday human language expressions), causing context regarding how the previous agent made decisions to be lost with each handoff. Ho explained that 'if 10 agents pass batons to each other, errors accumulate with each pass,' illustrating how errors accumulated structurally in the system.

Based on this diagnosis, the chosen new design was 'skill-tool architecture.' Rather than arranging independent agents as before, this approach decomposes functionality into smaller 'skills' and 'tools' components that can be shared, then combines them to construct the system. Ho noted that in achieving the reconstruction within 60 days, internal persuasion proved more difficult than technical challenges.

Persuading executives employed a demonstrative approach: creating a demo of the new design using actual customer inquiry data and comparing it side-by-side with existing system performance on the same tasks. Ho stated, 'At least in my view, the best evidence is what customers are trying to do. Whatever system we build must address that problem.' However, persuading the engineers who actually built the specialized agents required different logic.

The rationale Ho presented to engineers was 'scale'—the potential for broader impact. While a single independent agent solves only a narrow range of problems, the common skills and tools embedded in the new design could contribute cross-functionally to all customers requiring those capabilities. In other words, what they had created wasn't being 'discarded' but rather being reborn in a form with wider influence, a message that apparently led to acceptance on the ground.

Intuit's case demonstrates how practical deployment of AI agents is a continuous series of trial and error. Agent technology remains in development, and 'which design is correct' remains largely unknowable until actually tested in production environments. This suggests that in this field, the speed of decision-making in early problem diagnosis and reconstruction is directly linked to competitive advantage, rather than defending a completed design.