Multi-agent systems are having a moment. They promise better decomposition, specialization, and flexibility than single-agent workflows, and in demos they often look dramatically more capable. But in production, many teams discover an uncomfortable truth: the gains come with an “agent tax.”

That tax shows up in places teams underestimate at the start—latency compounding across chained calls, brittle handoffs between agents, state and memory drift, tool failures that cascade across the workflow, higher evaluation complexity, and human escalation paths that were never fully designed. What looks elegant on a whiteboard can become expensive, opaque, and hard to debug in the real world.

This session shares practical lessons from building and operating agentic systems beyond the prototype stage. I’ll walk through the hidden costs that emerge in multi-agent architectures, when those costs are justified, and the engineering patterns that make these systems manageable: constrained orchestration, clear ownership of state, failure isolation, observability for intermediate steps, and selective human-in-the-loop design. Attendees will leave with a sharper framework for deciding when multi-agent systems are worth the complexity—and how to build them so they survive production.

3 learning outcomes

1. Recognize the real “agent tax”: latency, coordination overhead, evaluation complexity, reliability risks, and operational burden that appear when multi-agent systems move into production.

2.Know when multi-agent design is actually justified versus when a simpler single-agent or deterministic workflow is the better system choice.

3. Apply practical operating patterns for production agentic systems, including orchestration boundaries, state management, observability, fault isolation, and human escalation design.

Speaker pitch / why me
I work on production-scale AI systems and have seen the gap between impressive agent demos and the realities of shipping dependable systems. My focus is on turning ambitious AI workflows into architectures that are observable, resilient, and operationally sane. This talk is aimed at engineering leaders and practitioners who want a grounded view of what multi-agent systems cost, where they create real value, and how to avoid the most common design mistakes.