Intuit trades multi-agent AI for skill-and-tool architecture
Intuit is moving away from a multi-agent architecture in favor of a granular skill-and-tool design, according to a VentureBeat preview of the company's VB Transform 2026 talk. Nhung Ho, Intuit's VP of AI, told the publication the company abandoned its "multi-agent system where we had large agents that did a lot" in favor of a model that decomposes agents into specialized components. The pivot matters because the source frames the change as a response to "customer expectations" shifting toward complex agentic tasks, while leaving out the production metric, cost structure, or failure mode that triggered the rebuild.
The new architecture, as Ho describes it, replaces large composite agents with a stack of "workflows, skills and tools down to the base level," paired with a new orchestrator and planner, and embeds human experts directly into the workflow alongside the AI. Intuit frames this redesign as "separating the brain from the hands," with the orchestrator and planner handling the coordination Ho calls "the brain" and individual skills executing the discrete actions a user task requires. Ho's quote acknowledges the scope of the change goes beyond the agent layer: "we changed the orchestrator, we changed the planner, we changed the brain, and we also changed what everybody had to build across the whole company." That last phrase is the one that determines whether the rebuild was strategic or forced, because rebuilding a system that every internal team builds against is rarely cheap, and the source does not address the cost or timeline of the migration.
Ho also describes the new architecture as letting Intuit "remain agile and use the best tools for the job, whether from large model providers or their own home-grown tools." The source presents this decoupling as an already-realized capability. Based on the design Ho describes, the planner-and-skills split is one of the patterns that does make provider substitution more tractable, because the skill layer can be written against a model-agnostic interface and the planner can choose between models per task. The source does not specify whether model substitution has actually happened in production, or whether the decoupling is theoretical and the system is still anchored to a single model.
Multi-agent orchestration continues to be the default pattern in vendor and consultancy pitches for complex, multi-step tasks. Intuit's pivot is one of the few public statements from a large enterprise saying that pattern did not survive contact with their workload. The source does not name what specifically broke: latency, cost, debugging burden, reliability, or the limits of context windows when agents chain together. That silence is consistent with a conference preview meant to set up the on-stage talk, but it leaves the trigger condition outside the article.
The preview also functions as a VB Transform 2026 promotion, with the publication listing other sessions on agentic orchestration from Target, Instacart, Asana, Rivian, Databricks, and Atlassian. None of those sessions include a comparable reversal narrative, and the source does not describe their architectures in enough detail to support cross-company comparison. The result is a preview that frames Intuit's rebuild as a confident advance, with the underlying technical and operational evidence deferred to the on-stage session. Whether the skill-and-tool design genuinely beats the multi-agent pattern on the specific production constraint that triggered the rebuild, or simply moves the cost somewhere else in the system, is the comparison the on-stage session at VB Transform 2026 on July 14 and 15 will need to deliver. The preview does not supply the evidence.