A single LLM call can answer a simple question, but complex tasks — researching a topic across ten sources, extracting and comparing product data from multiple competitors, or filling out a web form based on retrieved information — require multiple steps and decisions. Agent orchestration frameworks (LangGraph, CrewAI, Anthropic's Agent SDK) chain LLM calls with tool use and branching logic to accomplish these tasks.
Orchestration patterns include: pipeline (sequential steps), router (a coordinator agent dispatches tasks to specialised sub-agents), and swarm (agents communicate peer-to-peer). Each pattern involves managing state (what has been done, what the results were), handling errors (retries, fallback agents), and synthesising outputs from multiple agents into a final result.
Web scraping is a foundational capability in agentic workflows. An orchestration framework that needs live data routes a sub-task to a scraping agent — using a tool like AlterLab's MCP server — and passes the structured result to the next agent in the pipeline.