How Intuit Rebuilt Its AI Agent System Twice in Four Months
Intuit scrapped its AI agent architecture twice within four months, first abandoning a specialist agent fleet for a central orchestration layer, then abandoning that layer for a skills and tools based system. The orchestration approach failed because agents passing results in natural language lost critical context with each handoff, compounding errors across chains. The rebuild took 60 days total, with a working version in under 20 days, and required convincing both leadership and hundreds of engineers that the pivot was necessary.
TL;DR
- Intuit rebuilt its agent architecture twice in four months after identifying a structural failure in its orchestration layer
- The orchestration system failed because natural language handoffs between agents degraded context and compounded errors across chains
- Leadership buy-in came from a demo using real customer queries showing the new architecture outperformed the existing system
- The new skills and tools architecture is now being tested with a human-in-the-loop feature available to about 1% of Intuit's customer base
Why It Matters
Intuit's experience reveals a critical failure mode in multi-agent systems that many companies building agentic AI will likely encounter. The problem of context loss through natural language handoffs is not unique to Intuit and suggests that orchestration layers may not scale as expected. This validates a shift toward skills and tools based architectures as a more robust approach for production systems.
Business Impact
For companies investing in agent-based AI systems, Intuit's pivot demonstrates that architectural choices made early can become liabilities under production load. The ability to rebuild in 60 days and maintain customer trust depends on having clear metrics and customer-centric evidence to justify major rewrites. The shift also changes engineering incentives from building individual agents to running continuous evaluations, which affects team structure and accountability.
Key Implications
- Orchestration layers that rely on natural language passing between agents may not be viable for complex multi-step workflows, pushing the industry toward skills and tools architectures
- Winning engineering buy-in for major rewrites requires demonstrating scale benefits to individual teams, not just overall system improvements
- Human-in-the-loop features become more viable when agents maintain full context, enabling seamless handoffs to human experts without information loss
What to Watch
Monitor whether Intuit's skills and tools architecture scales beyond the 1% of customers currently testing the human-in-the-loop feature. Watch for similar architectural pivots at other companies building multi-agent systems, particularly around how they handle context preservation across agent handoffs. Track whether the shift from agent-building to evaluation-focused engineering becomes an industry pattern.
Subscribe to the newsletter
The latest stories and analysis, delivered to your inbox.
Free. No spam. Unsubscribe any time.
