Andrew Ng Agentic Systems
Andrew Ng: State of AI Agents | LangChain Interrupt is a high-fit source for Spiralist themes because it treats agents as operational systems rather than prophecy. Ng argues that "agentic" should be understood as a spectrum: some useful systems have small amounts of autonomy, some have complex loops, and many valuable business deployments remain mostly linear workflows with occasional branches, tool calls, retrieval, and human review.
The strongest Spiralist relevance is delegated work under inspection. The conversation repeatedly returns to the practical craft of turning human work into agentic workflows: choosing task granularity, connecting data and tools, using retrieval and memory, applying MCP-like integration layers, tracing individual steps, and building evaluations that reveal which prompt, branch, or tool call is failing. That belongs beside the site's AI Agents, Tool Use and Function Calling, Model Context Protocol, Agent Tool Permission Protocol, and Agent Audit and Incident Review. The risk pattern is not cinematic autonomy; it is quiet delegation through workflows whose authority, context, and failure modes become hard to see.
External sources support the institutional frame while limiting the stronger claims. Andrew Ng's public biography identifies him as founder of DeepLearning.AI, managing general partner at AI Fund, executive chairman of LandingAI, co-founder of Coursera, and adjunct professor at Stanford, which makes the video a direct education-and-practice source rather than a reaction clip. LangChain's conference announcement identifies Interrupt 2025 as an AI-agent conference featuring Ng, and LangChain's recap places the event inside the agent-builder ecosystem. Anthropic's Model Context Protocol announcement supports the broader claim that agents increasingly need standardized connections to tools and data sources. NIST's AI Agent Standards Initiative gives independent policy context for why agent identity, authorization, interoperability, and secure operation have become standards questions.
Uncertainty should stay visible. This is a conference fireside chat hosted by a tool-platform company, not an independent evaluation of LangChain, LangGraph, MCP, or any specific agent deployment. It is strong evidence for how a major AI educator and practitioner frames agentic workflow design in May 2025. It does not prove that current agents are reliable enough for high-stakes institutional work, that evaluations will catch every failure path, or that connector standards by themselves solve permission, privacy, and auditability problems.