Agentic workflow
A workflow where an LLM doesn't just answer once, but plans, uses tools, evaluates its own results, and iterates toward a goal over several steps. The shift from "a chatbot that answers" to "an agent that performs tasks".
Patterns (popularized by Andrew Ng and others): reflection (the model reviews and improves its own draft), tool use (search, run code, call APIs), planning (break a task into sub-steps), and multi-agent (several specialized agents collaborate). Instead of a single call it becomes a loop or a graph of calls driving toward a goal, often with ReAct as the core engine. Win: it can solve complex, multi-step tasks (code a whole function with tests, research and write a report) that a single answer can't. Price: more calls (cost/latency), harder to debug, and a risk that the agent "goes off the rails". The core of the 2024–2025 agent wave. Related to ReAct agent and multi-agent system.