Andrew Ng released this roughly one-hour free course in July 2026 on building agentic knowledge graphs from scratch. It works through three layers: what agentic knowledge graphs are, how to build your first graph, and how multi-agent systems are architected on top of graph structures — with hands-on code demos using Google ADK (Agent Development Kit). Our original coverage: Andrew Ng releases free course on building agentic knowledge graphs.
Per the publicly shared course timeline:
The course positions the agentic knowledge graph as a foundational component of agent systems: graph structure provides information persistence and relationship modeling, so agents do not start from zero on every task. The build path starts with simple nodes and edges and scales up to architectures supporting multi-agent collaboration — the graph serves as the memory and reasoning substrate.
This is where the phrase "graph engineering for multi-agentic systems" comes from: when multiple agents need to share state, pass context and coordinate work, a graph offers far more expressive power than a flat message queue.
The course fits developers who can already write basic agent logic and want to solve the "agents have no long-term memory" problem. At one hour, it is a conceptual introduction plus one complete walkthrough, not a systematic curriculum — you get the most value by reproducing the Google ADK segment (from 23:00) hands-on. For enterprises, the takeaway is evaluating whether existing agent systems need a graph component for long-term memory and coherence.
Where to watch: the course is distributed through Andrew Ng's official channels (DeepLearning.AI and his social accounts); check his official posts for the exact platform.
The course addresses agent memory, but production agents have two more hard prerequisites: code generation reliability and instruction compliance over long dialogues. The YZ Index continuously benchmarks both in public — the Execution dimension actually runs model-generated programs in isolated sandboxes, and the WDCD test measures whether constraints survive multi-turn pressure. When choosing the base model powering an agent system, these two dimensions belong next to capability scores: see the current leaderboard.
Yes. The course runs about one hour, was released in July 2026, and is distributed through Andrew Ng's official channels (DeepLearning.AI and his social accounts).
It suits developers with basic programming skills and familiarity with LLM agent concepts. It starts from "what graphs are and why agents need them," so the entry bar is low, but the Google ADK hands-on segment requires practical coding.
ADK is the Agent Development Kit, Google's toolkit for agent development. From 23:00 the course uses it to demonstrate combining graph structures with agent logic — the most hands-on part of the course.
A traditional knowledge graph is a static entity-relationship store built mainly for retrieval. An agentic knowledge graph is built and updated by agents at runtime, serving as the memory and reasoning substrate of an agent system — the emphasis shifts from storing knowledge to sustaining persistent state and collaboration across multiple agents.
