There is a particular kind of honesty that only comes from admitting you built the wrong thing twice in a single season. When Intuit's VP of AI, Nhung Ho, stood at VB Transform 2026 and walked through the company's agent architecture saga, she wasn't just sharing a war story. She was validating something we've suspected about this industry: the fastest path forward is often the one that forces you to throw away your own work. Intuit's journey, from a fleet of specialist agents to a central orchestrator and then to a skills-and-tools system, is a masterclass in treating your own architecture like a hypothesis rather than a monument. That is a hard lesson for any engineering culture, and it is exactly the one most teams are going to need to learn as they wade into agentic AI.
The failure mode that doomed the orchestration layer is worth sitting with, because it is not exotic. Passing results between agents in natural language meant every handoff lost context, and a ten-agent chain compounded errors by design. This is the kind of structural flaw that looks fine on a whiteboard and then quietly corrodes a production system. The related challenge of Clean Data Starts With Catching AI Slop Before It Skews Your Model is similar in spirit: garbage in, garbage out, whether the garbage is bad training data or a bad handoff protocol. And just as Transparency in AI Voice: ElevenLabs CEO on Disclosure and the Future forces us to ask who is responsible when an AI speaks, Intuit's rebuild forces us to ask who is responsible when an AI misroutes a financial task. The answer, for Intuit, was to stop pretending a central brain could manage everything and instead break the work into discrete, measurable pieces.
What makes this story more than a technical post-mortem is the human element. Ho's team had to convince hundreds of engineers who built the original agents to dismantle their own work. That is not a technical problem; it is a leadership problem. The winning argument was scale: a standalone agent solves one narrow problem, but a shared skill or tool can serve every customer who touches that part of the product. We would tell any reader staring down a similar rebuild to steal that playbook. Do not argue about architecture in the abstract. Build a demo on real customer queries, show it beating the existing system on the same tasks, and let the evidence carry the meeting. That is how you get buy-in without burning the building down.
The most forward-looking detail here is the shift in feedback loops. Moving from 0.3% of customers ever giving explicit feedback to nearly 100% because every chat is a feedback signal is not an incremental improvement. It is a different animal. Customers will tell your agent exactly where it failed, often in blunt terms, but they are also willing to give the system grace if it corrects course. The onus is on the builders to harvest that signal systematically. The practical takeaway for any team is this: if you are not designing for high-volume, plain-spoken feedback, you are flying blind. And if you are not prepared to scrap your own orchestration layer within four months when the evidence says it is compounding errors, you are not building agentic AI, you are just building a more complicated way to fail. Watch how Intuit scales that human handoff feature beyond the 1% test group, because that is where the real stress test begins.