An agentic workflow is a way of using a large language model in which, instead of asking the model to produce its final answer in a single pass, the system prompts the model repeatedly so it can build toward a higher-quality result step by step. The model might draft an answer, critique it, gather information with a tool, revise, and only then deliver an output. Andrew Ng popularized this framing in his DeepLearning.AI newsletter The Batch, beginning with the issue dated March 27, 2024.
Ng grouped the techniques into four design patterns he expected to drive much of the year’s progress. Reflection has the model criticize and improve its own output, optionally by pairing one agent that generates with another that critiques. Tool use connects the model to databases, APIs, and external services so it can act rather than only write. Planning breaks a complex task into executable steps that the model can follow and adapt when things go wrong. Multi-agent collaboration coordinates several specialized agents, each handling part of the work.
The framing mattered less as a new algorithm than as a vocabulary. By naming the patterns, Ng gave practitioners a shared way to describe what frameworks like ReAct, Reflexion, AutoGen, and CrewAI were each doing, and a checklist for designing agents. His core empirical claim was that a weaker base model wrapped in a good agentic workflow could outperform a stronger model used in a single shot.
Why business readers should care: “agentic” became a marketing label attached to almost anything in 2024 and 2025. The four-pattern framing is a practical filter - ask which of reflection, tool use, planning, and multi-agent collaboration a product actually implements, and whether the extra model calls those patterns require are worth the added latency and cost for the task at hand.