AI
Agentic Workflows
AI systems that plan, act, and iterate
Overview
Agentic AI workflows move beyond single-turn question answering. An AI agent perceives its environment, plans a sequence of actions, invokes tools (web search, code execution, file I/O, APIs), evaluates its own outputs, and iterates until a goal is achieved. Frameworks like LangChain, LangGraph, and AutoGen provide orchestration.
Key Concepts
- Task planning: decomposing goals into sub-tasks
- Tool use: calling external APIs, search engines, or code interpreters
- Memory: short-term (context window) and long-term (vector store)
- Self-reflection: evaluating and correcting intermediate outputs
- Multi-agent coordination: specialist agents collaborating on complex tasks
Key Facts
- Agentic systems introduce "tool injection" and "prompt injection" security risks
- Anthropic's Claude, OpenAI GPT-4, and Gemini all support function/tool calling
- ReAct (Reason + Act) combines chain-of-thought with tool use
- Human-in-the-loop remains essential for high-stakes agent decisions