Intuit rebuilt its AI agent architecture twice in roughly four months, Nhung Ho, the company’s VP of AI, disclosed at VB Transform 2026. The rapid succession of rebuilds reflected persistent reliability problems that made the previous systems unsuitable for production use.
The original architecture relied on a fleet of specialist agents, each handling distinct tasks within Intuit‘s business platform. A central orchestration layer was later introduced to coordinate these agents and streamline task routing. Agents communicated primarily through natural-language messages, and large, monolithic units combined planning and execution within a single component. The approach prioritized broad workflow coverage but introduced structural weaknesses that compounded over time.
Large, monolithic agents combining planning and execution introduced structural weaknesses that compounded over time.
Natural-language handoffs between agents caused context loss, leaving downstream agents without key details from earlier steps. That context drift produced compounding errors across workflows and generated error chains that were difficult to trace back to their origin. The coordination overhead also increased latency, reducing responsiveness for fast or complex tasks. These failure modes were significant enough to prompt a full architectural overhaul rather than incremental fixes.
The first rebuild replaced the specialist-agent fleet with a central orchestration layer. When that approach failed to resolve the underlying reliability issues, Intuit undertook a second rebuild, this time shifting to a skills-and-tools-based architecture. The second rebuild took 60 days in total, with a first working version delivered in under 20 days. Ho described the cycle of rapid rebuilds as the fastest path to success, despite the substantial restructuring each iteration required.
The new architecture decomposes large agents into smaller, specialized components, separating planning functions from execution functions. Skills encapsulate narrow, well-defined capabilities, while tools provide concrete actions or external integrations. Together, they replace the loosely structured natural-language handoffs of the prior system with structured calls and explicit function interfaces.
The design minimizes coordination overhead and tightens control loops, making the system more suitable for fast and complex tasks. By breaking down what were previously combined “brain and hands” agents into distinct planning and execution layers, Intuit gained finer control over behavior chains and reduced the surface area for cascading failures. The shift also made individual components easier to inspect and debug, addressing one of the core operational problems that plagued the multi-agent setup.
The rebuild also incorporated human experts directly into workflows alongside AI components, creating a hybrid human-AI system. Embedding humans within the architecture rather than positioning them only at the edges gives the system additional checkpoints for catching errors that automated components might propagate unchecked. A new human-in-the-loop feature currently in early testing allows customers to connect with an Intuit support representative, their accountant, or a bookkeeper mid-conversation while the system maintains full context throughout.
The two rebuilds underscore how quickly AI agent architectures can accumulate structural debt when deployed at production scale. For Intuit, the cost of scrapping and rebuilding proved lower than the cost of operating a system whose reliability could not meet the demands of its business platform.
