Intuit rebuilt AI agent system twice in 4 months, scrapped orchestration layer

Intuit began as an early pioneer in agentic AI, building systems where multiple specialist agents could collaborate to complete tasks. However, the company's journey to a working architecture has involved two significant redesigns in rapid succession.

The first architecture relied on specialist agents—each optimized for a narrow domain or task. Within four months, Intuit's engineering team recognized this approach had limits and rebuilt the system around a central orchestration layer, a coordinator agent that would delegate work to specialists and manage the flow of information between them. This seemed like a natural evolution: one intelligent agent routing tasks to others.

But the orchestration layer itself became the problem. As agents in the orchestrated system passed results to each other, each handoff was conducted in natural language. Nhung Ho, Intuit's VP of AI, explained the failure at VB Transform 2026: "If you have 10 agents and they all are passing to each other, every time that pass happens, error compounds." Each intermediate agent received context from the previous one, but natural language communication is inherently lossy—nuance and critical details were dropped, and downstream agents lacked the full picture they needed to act correctly.

Faced with this architectural dead-end, Intuit made the decision to rebuild entirely. The second redesign abandoned the orchestration layer in favor of a skills and tools based system—one where agents access tools and skills directly rather than routing through a central coordinator. This approach eliminates the intermediate handoff bottleneck. The rebuild was executed rapidly: the full redesign took 60 days, with a first working version ready in under 20 days. Ho described this pace of rebuilding as the "fast path" to success, suggesting that identifying and pivoting away from a broken architecture is faster than attempting to patch it.

Intuit's experience underscores a critical tension in agentic AI: the intuitive appeal of a central orchestrator (a single coordinator managing specialist agents) collapses under the weight of its own coordination overhead. The failure mode Ho described—context loss during natural-language handoffs between agents—is not a coding bug but a structural flaw: every agent in a chain depends on receiving complete, accurate state from its predecessor, yet natural language is lossy and ambiguous by design.

The move to a skills and tools based system suggests Intuit concluded that avoiding the orchestration layer altogether is preferable to patching it. Rather than having a master agent coordinate 10 specialist agents (each introducing a potential error point), the new model likely gives agents direct access to tools and skills without intermediate translation steps. This trades the simplicity of a unified orchestrator for the complexity of each agent being able to call the right tool at the right time—but eliminates the context-loss bottleneck. Ho's framing of two rebuilds in four months as "the fast path" implies the team views rapid experimentation and architecture-level pivots as faster than incremental tuning of a broken model.

No comments yet. Be the first to share your thoughts!