Chase Roossin and Steven Kulesza of Intuit Discuss the Hardest Challenge in Engineering: Coordinating AI Agents in Complex Systems

When Chase Roossin and Steven Kulesza from Intuit sat down with the Stack Overflow Podcast in April 2026 to discuss the hardest problem in engineering—getting multiple AI agents to function together in complex systems—they weren’t just talking about abstract algorithms. They were describing a challenge that’s quietly reshaping how cities like Austin, Texas, build and maintain the digital infrastructure powering everything from traffic management to utility grids. As someone who’s spent years covering the intersection of technology and urban development, I’ve seen how these multi-agent systems aren’t just theoretical—they’re already humming beneath the surface of places like Austin, where rapid growth collides with legacy systems, creating pressure points that demand smarter coordination.

Their conversation centered on what they called making agents “play nice at scale”—a phrase that sticks given that it captures the human-like friction in machine collaboration. Roossin, as group engineering manager, emphasized that success isn’t just about building smarter individual agents. it’s about designing the rules of engagement so they don’t step on each other’s toes when deployed in fleets. Kulesza, the staff software engineer, pointed to automated evaluations as the linchpin: rigorous testing frameworks that simulate edge cases before agents ever touch live data. This isn’t just about preventing errors—it’s about creating predictability in systems where uncertainty can cascade into real-world consequences, like a misrouted emergency response or a stalled public transit update.

What makes this especially relevant in Austin is how the city’s tech ecosystem has evolved. Home to major campuses for companies like Intel, Apple, and indeed Intuit’s own growing presence in the Domain Northside, Austin has become a laboratory for scaling AI systems. The University of Texas at Austin’s Oden Institute for Computational Engineering and Sciences has been researching multi-agent coordination for years, particularly in the context of smart grid resilience. Meanwhile, the City of Austin’s Innovation Office has piloted agent-based models for managing water distribution during droughts, where dozens of autonomous sensors and valves must negotiate flow rates without central oversight. These aren’t futuristic pilots—they’re live systems where the lessons from Roossin and Kulesza’s work on predictability and architectural trade-offs are being tested daily.

They contrasted two approaches: agent swarms—groups of simpler, diverse agents working in concert—and single, highly skilled agents attempting to handle everything. In Austin’s context, the swarm model often makes more sense. Think of the Austin-Bergstrom International Airport’s ground operations: instead of one omnipotent AI directing every baggage cart, fuel truck, and gate agent, a swarm of specialized agents handles specific tasks—one optimizing tug routes, another monitoring ramp congestion, a third adjusting to weather delays—all coordinated through shared protocols. This mirrors what Intuit found: swarms leverage redundancy and adaptability, which is crucial when dealing with the unpredictable rhythms of a growing city. But as Kulesza noted, this only works if you’ve built in automated evaluations to catch the strange edge cases—like how a sudden thunderstorm might cause agents to misinterpret runway availability.

Their discussion also highlighted how customer behavior shapes technical architecture—a detail that resonates deeply with Austin’s unique blend of tech workers, musicians, and entrepreneurs. Intuit’s teams apply real-world usage data to decide where to invest in reusability, ensuring AI components serve actual needs rather than engineering elegance. In Austin, this translates to how the Capital Metropolitan Transportation Authority (CapMetro) is refining its AI-driven bus scheduling system. Rather than imposing a rigid algorithm, they’re observing how riders actually transfer between routes during South by Southwest or ACL Fest, then using those patterns to train agent swarms that adapt to surges without overloading downtown corridors. It’s a practical application of the podcast’s core insight: architecture must follow behavior, not the other way around.

Of course, scaling these systems isn’t just a technical hurdle—it’s a socio-economic one. As Austin grapples with affordability crises and equity gaps, the deployment of AI agents risks amplifying disparities if not carefully managed. Roossin hinted at this when discussing evaluation frameworks: predictability isn’t just about system uptime; it’s about ensuring that an agent allocating public housing vouchers or optimizing food bank delivery routes doesn’t inadvertently disadvantage certain neighborhoods. The Austin Justice Coalition has been vocal about needing algorithmic transparency in municipal AI, and the principles from this podcast—rigorous evals, swarm diversity, customer-driven design—offer a pathway to audit and improve these systems before they’re baked into citywide infrastructure.

Given my background in urban technology policy, if this trend impacts you in Austin, here are the three types of local professionals you need to know about when navigating multi-agent AI systems in your organization or community project:

• AI Systems Architects Specializing in Municipal Infrastructure: Look for professionals with proven experience designing agent-based systems for public utilities or transit agencies, ideally with project involvement at CapMetro, Austin Water, or the Austin Energy grid. They should demonstrate familiarity with automated evaluation pipelines and be able to show how they’ve tested agent swarms against real-world scenarios like extreme weather events or festival-induced demand spikes. Question for case studies where they improved predictability without increasing computational overhead.

• Ethical AI Auditors with Local Government Experience: Seek experts who’ve worked directly with the City of Austin’s Innovation Office or the Office of Police Oversight on algorithmic impact assessments. Their criteria should include knowledge of Austin’s Equity Action Plan and ability to evaluate whether agent swarms perpetuate bias in resource allocation—whether that’s in emergency response routing or affordable housing waitlists. They’ll help you implement evaluation metrics that measure fairness alongside performance.

• Applied Research Liaisons from UT Austin’s Oden Institute: These aren’t just academics—they’re practitioners who bridge theoretical multi-agent theory with city-scale deployment. Prioritize those who’ve collaborated with the Texas Advanced Computing Center (TACC) on projects involving disaster response simulations or urban mobility modeling. They can help you design automated eval frameworks that leverage TACC’s computing power to stress-test your agent systems against Austin-specific variables, from I-35 traffic patterns to Barton Springs watershed dynamics.

Ready to find trusted professionals? Browse our complete directory of top-rated podcast,se-tech,se-stackoverflow,ai experts in the Austin area today.