Intuit didn’t get its AI agent architecture right on the first try. Or the second. At VB Transform 2026, the company’s VP of AI, Nhung Ho, told attendees that Intuit rebuilt its entire agent system architecture twice inside a four-month window, first replacing a fleet of specialist agents with a central orchestration layer, then replacing that too when the orchestrator started collapsing under its own complexity.

Ho’s framing was the part that stuck: scrapping recent work wasn’t a failure. It was the fast path.

That’s a hard thing to say in a company that had just built something. But it’s also the kind of honest reckoning with agent complexity that is becoming more common as enterprises move from AI pilots into production.

What Actually Happened at Intuit

The first redesign made sense on paper. Intuit had a growing fleet of specialist agents, each handling a narrow task. The problem was coordination. Without a central layer to direct traffic and synthesize outputs, agents couldn’t work together effectively. So Intuit built an orchestration layer to sit on top.

That orchestrator broke.

The failure mode was subtle. Because agents were passing outcomes to each other in natural language, each downstream agent had to infer how the upstream agent reached its conclusion. That inference degraded with every additional hop. A ten-agent chain didn’t fail occasionally; it compounded errors by design. The more capable the system tried to become, the more unstable it got.

That diagnosis led to the second rebuild: a skills and tools-based architecture, where agents aren’t handing off inferred context but calling defined, structured operations. It took 60 days to fully deploy, with the first working version running in under 20.

The Broader Problem

Intuit is not an outlier. At the same conference, Meta’s infrastructure VP Barak Yagour noted that agentic queries hitting Meta’s data systems grew 30x in a single six-month window, an inversion that is straining infrastructure built around human-paced, deterministic requests. His estimate: enterprises have roughly 20 months to fundamentally rebuild their systems to support agentic AI, or risk being caught flat-footed by the shift.

The pattern across both talks points to the same underlying reality. Agent infrastructure isn’t just a software problem. It’s an architectural one. The choices made early, about how agents communicate, how context is passed, how orchestration is structured, determine whether a system can scale or whether it becomes a liability.

Most companies currently deploying AI agents haven’t confronted this yet. Many are running what VentureBeat’s own research describes as “chatbot wrappers,” systems that look agentic in demos but don’t actually execute multi-step workflows autonomously. The gap between that and a production agent system is significant, and it’s where rebuilds like Intuit’s happen.

What This Means for Business

If a company as sophisticated as Intuit had to scrap and rebuild twice, what does that say about enterprises in general?

A few things.

Architecture choices compound. The design decision you make about how your agents communicate today will determine how much rework you’re doing in six months. Natural language handoffs between agents seem flexible, but they introduce inference errors that compound at scale. Structured tools and skills take longer to define upfront but don’t degrade under load.

Iteration speed matters more than getting it right the first time. Nhung Ho’s framing, that two rebuilds in four months was the fast path, is worth taking seriously. The companies that try to perfect their architecture before shipping are the ones that get passed by. The ones that can rebuild quickly when something isn’t working are the ones that end up ahead.

You need the organizational capacity to scrap work. Ho was explicit that the harder problem was internal, convincing leadership and engineers who had built the original agents that the right call was to throw it away. That’s a culture question as much as a technical one. Companies without that culture tend to patch failing architectures rather than replace them.

The “wait until agents are more mature” strategy has a cost. Meta’s 20-month window isn’t a hard deadline, but it reflects something real: infrastructure built for human-paced, deterministic workflows is already straining under agentic load. Waiting to invest in agent infrastructure doesn’t eliminate the need to rebuild; it just means you’re doing it under more pressure.

What Experienced Operators Know

One of the consistent themes across VB Transform 2026 was the gap between how AI agents are marketed and how they actually behave in production. The companies making progress, Intuit included, are the ones who have stopped treating agent deployment as a product rollout and started treating it as infrastructure design.

That shift requires different skills than most businesses have internally. It requires understanding how agents communicate, how errors propagate through multi-step systems, and what kind of architecture can actually handle the failure modes that show up in production, not just in demos.

For businesses considering their first serious AI agent deployment, Intuit’s story is both reassuring and clarifying. Reassuring because even large, capable tech companies had to iterate significantly to get this right. Clarifying because it sets realistic expectations: building a real AI agent system isn’t something you get right the first time. The question is how quickly you can identify what’s not working and rebuild.

---

If you’re evaluating AI agent deployment for your business and want to avoid the most expensive architectural mistakes, book a discovery call with the Omni team. We work through the architectural and operational decisions that determine whether AI agent investments deliver or need to be rebuilt.