For most of 2025 and into 2026, the default answer to “how do we scale enterprise AI” was: add more agents. Spin up a planner agent, a research agent, a writer agent, a reviewer agent, wire them into a multi-agent system, and call it orchestration. Intuit tried that too. Then it tore the design apart and rebuilt around something narrower: a skill-based architecture where tools are coordinated by a single planner, with human experts kept in the loop on purpose, not as a fallback.
What Intuit actually changed
Nhung Ho, VP of AI at Intuit, put it plainly: “We went from a multi-agent system where we had large agents that did a lot to fully incorporating workflows, skills and tools down to the base level.” Her team didn’t tweak a prompt or swap a model. They changed the orchestrator, the planner, and what every team across the company had to build.
The rebuilt system has three layers. A capability layer lets new skills, lending, payroll, tax, get onboarded as modular components instead of new monolithic agents. A trust layer runs a continuous evaluation engine that scores every agent output for accuracy and flags anything under a confidence threshold. A scale layer means adding a capability is a configuration change, not a new codebase. The side effect worth noting: this also decouples Intuit’s orchestration from any single model vendor, so the company can swap models without rewriting the workflow logic around them.
Why big multi-agent systems kept breaking
Multi-agent systems look elegant in a demo. In production, they accumulate a specific kind of failure: context inconsistency across agents, not the orchestration pattern itself, is the most common reason these systems break down. One agent updates its understanding of a task; the others don’t. Left unmanaged, that mismatch cascades. A stalled agent doesn’t just stall, it can spiral into feedback loops, generate false consensus with a peer agent, and burn through an API budget in minutes.
Large, broad-capability agents make this worse because each one is doing several jobs at once: planning, reasoning, calling tools, and judging its own output. When something goes wrong, you can’t easily tell which of those four jobs failed. Splitting a big agent into discrete skills behind one planner narrows the blast radius. If the tax-filing skill misfires, the planner can catch it, retry, or route to a human, without the payroll skill ever knowing there was a problem.
The infrastructure math backs this up
There’s a cost argument sitting next to the reliability one. A VentureBeat Research survey of 573 technical leaders at companies with 100 or more employees found that 86% of enterprises running their own GPUs report utilization at 50% or less. Most of that capacity is already paid for and sitting idle.
Stack a sprawling multi-agent system on top of that and the math gets worse, not better. Every extra agent in a chain is another round of model calls, another chance for a retry loop, another line item on the compute bill. A skill-based design with one planner making fewer, more deliberate calls is a smaller number to explain to finance, and a smaller number of things that can go wrong at 2 a.m.
Humans, on purpose
The part of Intuit’s redesign that’s easy to miss: it puts human experts back into the workflow deliberately, at specific checkpoints, instead of trying to automate them out. The Payments Agent predicts a late invoice and drafts a reminder, but the business owner reviews and edits it before it goes out. The Finance Agent builds a forecast or a KPI dashboard, but a human accountant decides what to actually do with it.
That pattern lines up with where a lot of enterprise AI design is heading in 2026, partly for practical reasons and partly because of regulation. The EU AI Act’s human-oversight requirements for high-risk systems are pushing teams to design checkpoints in from the start rather than bolt them on later. Either way, the shape of a “trustworthy” agent system in 2026 looks less like full autonomy and more like a well-placed pause button.
What to change in your own agent architecture
If you’re deciding how to architect an AI system for production rather than a demo, Intuit’s rebuild suggests a few concrete moves:
- Break a large, do-everything agent into discrete skills and tools that sit behind a single planner, instead of chaining several broad agents together.
- Add an evaluation layer that scores output confidence and routes low-confidence results to a human, rather than letting a downstream agent consume unverified output.
- Design human checkpoints at the specific steps where a wrong call is expensive, like sending a payment reminder or filing a return, and let the agent run autonomously everywhere else.
- Check GPU and inference utilization before adding more agents to a workflow. If existing capacity is already running under 50%, the fix is rarely “more agents.”
Conclusion
None of this means multi-agent orchestration is a dead end. It means it’s a tool for a specific problem, not the default answer to every AI architecture question. Intuit had the scale and the failure data to justify tearing down a system it had already built and shipping something narrower. Most teams won’t get that same forcing function until something breaks in production. Better to ask the architecture question now: does this task actually need five agents talking to each other, or does it need one planner, three well-scoped skills, and a human who reviews the output before it goes out the door?