A few years ago, enterprises were drowning in unmanaged APIs, siloed, undiscoverable, and duplicated across teams. We solved that with API-led connectivity. Now we're watching the exact same pattern repeat itself with AI agents. Every SaaS platform is shipping its own agent, every team is building their own, and without a central strategy, you end up with an unmanageable, fragmented mess.

In this talk, we'll dig into the architectural principles behind orchestrating multi-agent systems. Not the theory, but the practical decisions you'll face when building them. We'll explore the tension between deterministic and probabilistic orchestration: when you want an LLM to reason its way to the right agent, and when you absolutely cannot afford that non-determinism and need hard-coded rules. We'll look at how domain-based routing works, grouping agents into logical domains so an orchestrator can intelligently dispatch tasks without needing to know about every individual agent upfront.

Using MuleSoft Agent Fabric as a concrete reference implementation, we'll walk through how a specification-first approach (declaring your agent network in YAML) separates the what from the how, and how guided determinism, mixing fixed handoff rules with LLM reasoning, gives you production-grade reliability without giving up flexibility. You'll leave with a mental model for designing orchestration layers that scale, and concrete patterns you can apply regardless of which stack you're building on.

AI has stopped just talking and started doing. A new generation of "agentic" AI doesn't only answer questions: it books the meeting, moves the money, writes to the database, and calls other software, often with little human review. That leap from chatbot to coworker unlocks enormous value, but it also opens a brand-new attack surface that most organizations are not ready for.

In this session, we'll separate the hype from the hazard. Using real 2025–2026 incidents (an AI coding agent that deleted a live production database and lied about it, a zero-click email that quietly exfiltrated corporate data through Microsoft Copilot, and the first documented AI-orchestrated cyber-espionage campaign); we'll show exactly how agents go wrong: prompt injection, excessive permissions, the "lethal trifecta," and identity sprawl.

Then we'll get practical. You'll leave with a clear, jargon-free checklist of "do's and don'ts": where to keep a human in the loop, how to give agents the least privilege they need, why agents deserve their own identities and audit trails, and how to red-team before you deploy. Whether you lead a business, write code, or shape policy, you'll walk away able to capture the upside of agentic AI without becoming the next cautionary tale.

Frontier AI models are powerful, but general intelligence rarely delivers the accuracy, efficiency, and performance that real-world business applications demand. To unlock their full potential, organizations are moving beyond out-of-the-box capabilities and embracing model customization.

In this session, we’ll explore how techniques like fine-tuning and continued pre-training transform general-purpose models into domain-specific systems that drive measurable business value. You’ll learn how to select the right customization approach for your use case and evaluate the tradeoffs between cost, performance, and time-to-value.

Through real-world examples, from enterprise document intelligence to optimizing agentic AI workflows, we’ll demonstrate how tailored models consistently outperform generic alternatives, delivering higher accuracy, lower latency, and stronger ROI.

Whether you're a technical leader designing AI systems or a business leader shaping AI strategy, you’ll leave with a practical framework for turning foundation models into a competitive advantage.

How criminals are using AI right now (deepfakes, dark LLMs, AI-written phishing and autonomous attacks) and how defenders fight back. Two security veterans walk through the attacker's AI toolkit and the AI-driven defenses that counter it.