From answering questions to completing tasks
Why 'agent' needs a better definition
Andrew Ng coined the term 'agentic AI workflows' because companies use 'agent' to mean wildly different things — from a single prompt to a complex multi-system pipeline. Calling it a workflow clarifies that it is a multi-step process, not a single magic box.
One-shot vs agentic
A one-shot RAG answer to 'what is your refund policy?' gives a generic policy quote. An agentic workflow retrieves the policy, asks for the order number, queries an API to check eligibility, and confirms the refund with a timeline. It completes the task, not just answers the question.
The workflow is the unit
The power is not in any single prompt or tool — it is in the orchestration. Each step handles one thing (extract, look up, check, draft, send), and the workflow chains them together. This is the foundation of every specialised AI system being built today.
Specialised agents are everywhere
AI software engineer, AI SDR (sales development rep), AI lawyer, AI skills mentor — these are all agentic workflows. They combine prompts, tools, and decision logic into a system that handles a specific job. Not all of them work well yet, but the pattern is clear.
Think about the last time you used a chatbot for customer support. What steps would a truly helpful agent need to take?