Agentic Conversations (formally mlops.community)

Sarah Drasner, Google Area Tech Lead and distinguished engineer, explores what happens when agents browse the web for us. She compares browser-use, computer-use, API, and MCP agents, then demos an agent gathering photos and logos for Canva while keeping creativity human. They unpack WebMCP’s origins, gateway possibilities, engineering instinct, thought experiments, and safer, more delightful agent experiences.

Justin Shroeder, co-founder of Standard Agents, explores a future powered by thousands of specialized AI agents instead of one all-purpose model. They discuss agent standards, runtimes, prompt engineering, and why narrow scope can improve reliability. The conversation tackles routine automation, managing countless agents, the possibility of agent marketplaces, and bringing AI builders together with domain experts.

Issac Lee, an AI and blockchain executive at NEXUS, explores MMORPGs populated by tireless AI players instead of scripted NPCs. He discusses adaptive companions, agent-generated worlds, and solving multiplayer launches with bots that never sleep. The conversation also visits NEXUS’s Da Vinci development system, text-to-3D tools, self-hosted AI infrastructure, personalized meeting notes, privacy, and shopping agents reshaping ecommerce.

Jason Ward, a C.H. Robinson FinOps practitioner and member of the FinOps Foundation’s AI working group, explores AI across logistics, from quoting to tracking. He explains secure model routing, tagging resources to control costs, and using AI to uncover observability signals. The conversation covers prompt bloat, context starvation, agentic versus conversational workloads, and experimental metrics like cost per thought.

James Barney, Global Head of AI Enablement at MetLife, explores the art of “caveman prompting” and the coming CaveBench challenge. They discuss matching models to workloads, balancing cost, speed, and accuracy, pricing surprise launches, measuring AI’s business value, and governing agents. The conversation also looks at whether chat interfaces could replace dashboards, and the risks of missing insights you never thought to ask for.

What comes after building an MCP server? This discussion explores production observability, smarter tool design, code mode, authentication, and the growing complexity of agent evaluations. It examines how AWS might tackle roughly 16,000 APIs, whether specialized servers beat unified access, and how prompt-injection warnings, token costs, and tool overload complicate real-world deployments.

Ola Hungerford, Principal Engineer for AI Enablement at Nordstrom and Model Context Protocol maintainer, explores why skills and MCP are complementary rather than interchangeable. They discuss turning tool collections into workflows, progressive discovery, distributing knowledge and memory, metadata risks, prompt injection, personas and sub-agents, and the challenge of standardizing fast-moving agent technology.

Angie Jones, VP of Developer Experience at the Agentic AI Foundation, brings field notes from China and Tokyo. She explores China’s super-app agent ecosystem, WebMCP’s page-level tools for logged-in browsing, and why computer-use agents need faster collaboration. The conversation also tackles agentic commerce, payment protocols, A2A services, privileged access, and the ongoing chaos of three-letter acronyms.

Abhinav Lad, Palo Alto Networks’ cloud and AI finance director, and Kuntal Patel, its cloud engineering and FinOps leader, explore why agentic AI broke traditional cost forecasting. They discuss runaway loops, smarter AI gateways, model trade-offs, retry limits, and measuring value through adoption, users, revenue, and engineering outcomes. They also explain why token pricing alone can send financial strategy in the wrong direction.

Ambud Sharma, a Pinterest Principal Engineer specializing in infrastructure economics and AI systems, presents a five-layer approach to scaling AI efficiently. They compare the radically different stacks behind products like Cursor and YouTube, then explore hardware planning, inference bottlenecks, model routing, quantization, governance, and the rising complexity of agent-to-agent workloads.