{"id":63892,"date":"2026-07-18T19:43:45","date_gmt":"2026-07-18T23:43:45","guid":{"rendered":"https:\/\/overcentral.com\/en\/?p=63892"},"modified":"2026-07-18T19:43:45","modified_gmt":"2026-07-18T23:43:45","slug":"intuit-agent-architecture-rebuild","status":"publish","type":"post","link":"https:\/\/overcentral.com\/en\/intuit-agent-architecture-rebuild\/","title":{"rendered":"Intuit scraps AI agent architecture twice in 4 months, calls it fast path"},"content":{"rendered":"

Intuit<\/a> was an early pioneer in the deployment of agentic AI<\/a> for automating complex financial tasks, but the company’s journey to a scalable architecture has been anything but linear. Speaking at VB Transform 2026, Intuit VP of AI<\/a> Nhung Ho detailed how the company scrapped and rebuilt its entire agent architecture twice in just four months, ultimately abandoning a central orchestration layer for a skills-and-tools-based system after the orchestrator itself became a bottleneck. The story of why those architectures failed, and what it took to rebuild them, offers a rare, candid look at the structural challenges facing any organization trying to scale agentic systems in production.<\/p>\n

Intuit’s agent architecture: from specialist fleets to a broken orchestrator<\/h2>\n

The original system was built around a fleet of specialist agents, each designed to handle a specific, narrow task. This approach solved one problem\u2014having capable agents\u2014but created another: the customer had to decide which agent to use for which job. Intuit\u2019s solution was to introduce a central orchestration layer that could take a natural language task and route it internally, removing the burden of choice from the user.<\/p>\n

That orchestration layer worked for roughly three months, a period Ho described, only half jokingly, as \u201ca year\u201d in the compressed development timeline of 2026. The breakdown was not due to capacity but to a fundamental structural flaw. The agents in the orchestrated system passed results to each other in natural language, and each handoff degraded the context the next agent needed to act correctly.<\/p>\n

\u201cIf you have 10 agents and they all are passing to each other, every time that pass happens, error compounds,\u201d Ho said.<\/p>\n

The result was not an occasional failure but a system that was designed to compound errors. Each downstream agent had to infer how the upstream agent had reached its conclusion, and that inference became progressively less reliable with every additional hop in the chain. A ten-agent pipeline did not fail sporadically; it was architecturally predisposed to produce degraded outputs.<\/p>\n

How Intuit rebuilt its agent system in 60 days<\/h2>\n

The diagnosis that the orchestration layer was broken by design sent Intuit back to a skills-and-tools architecture. What followed was a full production rebuild completed in 60 days, with a first working version ready in under 20. Ho\u2019s team demonstrated the new architecture using real customer queries, showing it outperforming the existing system on identical tasks.<\/p>\n

\u201cThe best proof, at least my belief, is what are customers trying to do? And whatever system you build needs to address those problems,\u201d Ho said.<\/p>\n

Rebuilding the system required more than a technical decision. Ho described the harder problem as internal: convincing both leadership and the engineering teams who had built the original agents that scrapping recent work was the correct strategic move. The pitch to leadership was grounded in evidence rather than argument, using live customer queries to illustrate performance gains.<\/p>\n

Winning over engineering required a different case. Hundreds of engineers outside Ho\u2019s core team had built the specialist agents that were being retired. The ask was to disassemble those agents into individual skills and tools that could be shared across the product. Ho framed the decision around scale: a standalone agent solved one narrow problem, while a shared skill or tool built into the new architecture could serve every customer who touched that part of the product. This shift fundamentally changed the day-to-day responsibilities of partner teams, moving their focus from building and maintaining agents to running evaluations, which became the only reliable way to measure whether the new architecture was working.<\/p>\n

Bringing humans into the loop: the customer-facing result<\/h2>\n

The clearest customer-facing outcome of the rebuild is a feature that allows a live agent conversation to pull in a human participant. Currently in early testing and live to about 1% of Intuit\u2019s customer base, the feature is scheduled for scaling in the coming weeks, according to Ho.<\/p>\n

In practice, a customer can bring an Intuit product support person, their own accountant, or one of Intuit\u2019s bookkeepers directly into the mid-conversation conversation. That human joins with the full context of everything the agent has already done. Ho directly contrasted this with how most AI chat products handle the same scenario: a general-purpose assistant answering a tax question typically ends with a disclaimer to consult a professional. Intuit\u2019s system is designed to connect the customer to that professional directly, inside the same conversation thread.<\/p>\n

The human handoff is paired with a permissions model built specifically for financial data. Every action an agent takes on a customer\u2019s financial data requires explicit permission upfront, though Ho noted that this requirement can be relaxed over time as customers build trust in the system. Intuit also maintains a complete audit log of every action an agent performs, and all actions are reversible if needed.<\/p>\n

Feedback in the agentic AI era: from sparse to total<\/h2>\n

The architecture rebuild also fundamentally changed how Intuit gathers and uses customer feedback\u2014a shift Ho described as qualitatively different from previous approaches.<\/p>\n

\u201cFeedback in the past used to be very, very sparse, and it was also very bimodal,\u201d Ho said. \u201cEither they loved it or they hated it, and usually it tends towards the negative.\u201d<\/p>\n

In a chat-based system, every conversation functions as a feedback signal. Ho said this moved Intuit from roughly 0.3% of customers ever giving explicit feedback to something close to 100%, producing a volume of data no manual review process could manage. Ho has returned to writing code herself specifically to build models that analyze that feedback systematically, looking for where the system falls short.<\/p>\n

That volume comes with a tone most product teams aren\u2019t used to hearing directly. Customers tell the agent exactly where it failed, in plain terms.<\/p>\n

\u201cThey straight up tell you, \u2018You suck. I hate this. This is not right,\u2019\u201d Ho said. \u201cBut they\u2019re also willing to give the systems grace and correct it as well, and so the onus is on all of us to<\/a> harvest this new piece of feedback and type of feedback, and actually improve the system.\u201d<\/p>\n

What this means for teams building agentic systems<\/h2>\n

Intuit\u2019s experience carries direct implications for any engineering organization building agentic AI into production. The failure of the central orchestration layer demonstrates that passing natural language results between agents creates a compounding error problem that grows with chain length. The move to skills and tools represents a different architectural philosophy: instead of routing tasks through a central dispatcher, the system draws from a shared library of composable capabilities.<\/p>\n

The rebuild\u2019s tight timeline\u201460 days for a full production system\u2014also shows that fast iteration is possible when the decision to scrap existing work is driven by clear evidence. Ho\u2019s team used real customer queries to make the case, which underscores a broader principle: in agentic AI, the most effective persuasive tool is a demonstration that the new system performs better on actual user tasks.<\/p>\n

The feedback transition from sparse to near-total is perhaps the most forward-looking implication. As conversational interfaces become the primary surface for agentic systems, product teams will suddenly have access to an unprecedented volume of direct, plain-language criticism. The organizations that succeed will be those that build the infrastructure to harvest that feedback, analyze it at scale, and close the loop by rapidly improving the system in response.<\/p>\n

For teams building similar systems, the immediate takeaway is to evaluate whether your current architecture compounds or contains errors across agent handoffs, and whether you have the evaluation infrastructure in place to measure performance on real customer tasks, not synthetic benchmarks.<\/p>\n","protected":false},"excerpt":{"rendered":"

Intuit was an early pioneer in the deployment of agentic AI for automating complex financial tasks, but the company’s journey to a scalable architecture has been anything but linear. Speaking at VB Transform 2026, Intuit VP of AI Nhung Ho detailed how the company scrapped and rebuilt its entire agent architecture twice in just four […]<\/p>\n","protected":false},"author":7,"featured_media":74681,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"fifu_image_url":"https:\/\/iili.io\/CjLg4iQ.jpg","fifu_image_alt":"Intuit scraps AI agent architecture twice in 4 months, calls it fast","footnotes":""},"categories":[349],"tags":[],"class_list":["post-63892","post","type-post","status-publish","format-standard","has-post-thumbnail","category-articles"],"fifu_image_url":"https:\/\/iili.io\/CjLg4iQ.jpg","fifu_image_alt":"Intuit scraps AI agent architecture twice in 4 months, calls it fast","_links":{"self":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/posts\/63892","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/users\/7"}],"replies":[{"embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/comments?post=63892"}],"version-history":[{"count":0,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/posts\/63892\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/media\/74681"}],"wp:attachment":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/media?parent=63892"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/categories?post=63892"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/tags?post=63892"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}