AI Superstream: Agent Orchestration

Schedule

The time frames are only estimates and may vary according to how the class is progressing.

Introduction – Vicki Reyzelman (5 minutes)

Vicki welcomes you to the AI Superstream.

Keynote: The Landscape of Agentic Orchestration – Matt McLarty (15 minutes)

AI agents bring exciting possibilities and revolutionary capabilities. But for organizations to realize value in the new world of orchestrated AI agent systems, they have to figure out how it fits with their current world. Boomi CTO Matt McLarty surveys the agentic landscape, summarizing the key primitives, patterns, and principles that will help you put agent orchestration into practice.

How to Govern, Supervise, and Continuously Improve Multi-Agent Systems – Tatyana Mamut (30 minutes)

Managing AI agents is more like managing humans than software, and constant supervision is required. When AI agents are orchestrated into multi-agent systems and workflows, the monitoring, supervision, and improvement cycles—reading all the logs, traces, and chain-of-thought reasoning blocks for all the agents, piecing them together, analyzing the full picture of the multi-agent application, assessing effectiveness of the end-to-end workflow, and producing suggestions for improvement—can become overwhelming. Tatyana Mamut, PhD and cofounder and CEO of Wayfound, takes you through the practical ways of ensuring that multi-agent systems stay compliant, effective, and continuously improving in production without a dedicated team of data scientists and engineers.

Break (5 minutes)

Context-Aware Foundations for Agentic AI – Tendu Yogurtcu (30 minutes)

Agentic AI systems depend on the quality, meaning, and governance of the data that informs their decisions. Context does not emerge automatically from model scale or orchestration logic. It's established through semantic clarity, metadata intelligence, and disciplined data architecture. As enterprises move from experimentation to production, fragmented metrics, inconsistent definitions, and weak governance become primary barriers to trust and scale. Reliable agentic systems and their orchestration layers are built on clearly defined business meaning, well-modeled data relationships, and policy enforced consistently across platforms. Join Tendu Yogurtcu to gain practical guidance for building production-ready agentic systems on strong data foundations, enabling scale, trust, and measurable enterprise value.

Orchestrating AI Agents When Physics Has the Final Word – Susanna Holt (30 minutes)

Most agent orchestration talks focus on efficiency: routing tasks, managing context, reducing cost. But what happens when your agents operate in a domain where physics is the final arbiter? Where a wrong answer isn’t just unhelpful; it’s physically impossible or unsafe. Susanna Holt, CTO at OLI, explores agent orchestration through the lens of OLI’s work on predicting chemical changes in water-based environments. You’ll examine the design patterns that emerge when you need a generative agent and a safety-focused validator agent to work together and examine the hard questions: How do agents negotiate disagreements when correctness isn’t optional? When should you replace an AI agent with deterministic computation? And how do you calibrate a system that must be cautious enough to be safe without being so cautious that it’s useless? The patterns are grounded in industrial chemistry, but they transfer to any domain where AI agents must respect hard constraints—from engineering and manufacturing to healthcare and finance.

Break (5 minutes)

Building Deterministic Guardrails for Autonomous Agents: Security for Autonomy, Tool Use, and Memory (ATM) – Sumeet Jeswani (30 minutes)

As organizations rapidly adopt agentic AI, traditional LLM security measures fall short. When building orchestration layers, developers are dealing with agents that make independent decisions, access third-party APIs, and retain long-term context. A failure in any of these areas can lead to catastrophic, systemic breaches across the enterprise stack. Drawing on principles from the OWASP Top 10 for Agentic Applications, Sumeet Jeswani, a senior security architect at Google Cloud, introduces the ATM framework. You’ll explore the critical difference between least privilege and least agency—why giving an agent access to a database is fundamentally different from giving it the autonomy to delete the database. Join Sumeet to learn how to identify hidden orchestration vulnerabilities and implement deterministic guardrails to prevent minor tool failures from escalating into enterprise-wide cascading failures.

What’s Old Is New Again: Going from Monoliths to Micro-Agent Architectures – Cornelia Davis (Sponsored by Temporal) (30 minutes)

In the 2010s, the shift from monoliths to microservices improved scalability, resilience, and team autonomy. Today, AI systems are undergoing a similar transition from large, general-purpose agents to modular, specialized micro-agents that offer greater flexibility and composability. However, this advance introduces coordination challenges. Cornelia Davis, principal developer advocate for Temporal, explores how to define agent boundaries and use durable orchestration patterns to build reliable, scalable multi-agent systems.

This session will be followed by a 30-minute Q&A in a breakout room. Stop by if you have more questions for Cornelia.

Break (5 minutes)

The Reliability Tax: Why Agentic AI Workflows Fail and What Architecture (Not Bigger Models) Can Do About It – Luis Sanchez (30 minutes)

A model with 95% per-step accuracy sounds production-ready—until you chain it across 100 agentic steps and its cumulative success collapses to 0.6%. This is the reliability tax, and it’s the reason most multi-agent systems never leave the proof-of-concept stage. Luis Sanchez, founder of Toryx AI and SGX Analytics, presents empirical findings from stress-testing frontier and open source models across real-world agentic workflows in insurance underwriting, financial risk analytics, and enterprise COBOL modernization. These are industries in which failure isn’t a demo glitch; it’s a regulatory event. Luis takes you through the failures he has cataloged: silent degradation, schema drift, hallucination cascades, context rot, and dialect-specific collapse in legacy code. He also shares a patented orchestration architecture to demonstrate how intelligent routing, validation-driven escalation, and multimodel consensus recover reliability at a fraction of single-model cost. You’ll leave with a framework for reasoning about failure in agentic systems, a catalog of the failure modes that kill production deployments, and an understanding of why the next breakthrough in agent orchestration is architectural risk management.

Architecting Agent Teams: Orchestration Patterns for Enterprise AI – Ayo Adedeji (30 minutes)

As AI moves from chat interfaces to autonomous workflows, the hard problem shifts from picking a model to orchestrating teams of agents in production. Ayo Adedeji, developer relations manager at NVIDIA, shows you how to architect specialized agent teams you can deploy, inspect, and measure. Using the open source NeMo Agent Toolkit as a unifying layer, Ayo demonstrates framework-agnostic orchestration: Connecting agents built in LangChain, CrewAI, Google ADK, and custom frameworks without replatforming. He then takes you through the NVIDIA AI-Q Blueprint as a reference architecture, with an orchestration node that classifies intent and routes between shallow and deep research agents.

Closing Remarks – Vicki Reyzelman (5 minutes)

Vicki closes out today’s event.