Andrew Ng on the State of AI Agents: From Agentic Systems to Underrated Tools

Andrew Ng on AI Agents

Andrew Ng discusses the current state of AI agents, emphasizing the concept of “agenticness” as a spectrum of autonomy rather than a binary state. He highlights the need for better skills in breaking down business processes into agentic workflows and the importance of robust evaluation frameworks. Ng also points out underrated areas like the voice stack and the necessity of widespread coding literacy in the age of AI.

Points clés

Andrew Ng has been a significant part of the LangChain journey, collaborating on six short courses with deeplearning.ai.
Ng advocates for viewing “agenticness” as a degree of autonomy in systems rather than strictly defining something as an agent.
He observes many business opportunities in automating linear workflows with occasional branches using agentic systems.
A key challenge in building agentic workflows is breaking down complex tasks into granular micro-tasks and implementing effective evaluation frameworks.
Ng suggests that many teams delay implementing systematic evaluations, relying too heavily on human assessment.
The voice stack is highlighted as an underrated area with significant potential for large enterprise applications.
AI-assisted coding is seen as a major productivity booster, and Ng believes more businesses need to adopt it.
Ng emphasizes the importance of coding literacy for everyone, not just software engineers, to better interact with computers.
MCP (Multi-Agent Communication Protocol) is viewed as a fantastic first step towards standardizing interfaces for tools and data sources, although it is still early in its development.
Ng notes that multi-agent systems built by a single team are more successful currently than interactions between agents developed by different teams.

À retenir

So, apparently, we shouldn’t waste time arguing if something is a true agent. Just call it “agentic” and move on. Also, if you’re not using AI to help you code, you’re basically living in the dark ages. And seriously, learn to code – even if you’re just answering phones or balancing books. It’ll apparently make you better at telling a computer what to do. Who knew?

Sources