AI Agent Orchestration Frameworks

AI Agent Orchestration Frameworks From Single Agent to Agent Teams

---

In this talk, we unpack the architecture behind autonomous AI agents and the emerging standards that let them work together at scale.
We start from first principles: what exactly is an agent? Beyond the LLM, every agent runs on a harness of memory, tools, and task-specific skills, and connects to the world through MCP, the protocol that's become the USB-C of AI. We then go deeper into how agents communicate with each other through the A2A protocol, forming teams that can discover, delegate, and collaborate across vendors and platforms.
From there, we examine some real-world agent frameworks that sit at opposite ends of the governance spectrum: OpenClaw, the open-source phenomenon that crossed 215,000 GitHub stars and sparked a mainstream debate about AI consent, and NemoClaw, NVIDIA's enterprise response that treats security as an architectural constraint, not an afterthought.
The talk closes with a practical map of the five orchestration patterns powering production agent systems today (Pipeline, Supervisor, Fan-out, Swarm, and Debate) and a survey of how the major labs (Anthropic, OpenAI, Google) have each built their own framework philosophy on top of them.
Whether you're building with agents, governing them, or just trying to understand what the industry is actually shipping in 2026, this talk gives you the vocabulary, the architecture, and the hard questions worth asking.