I Split One AI Agent Into Two - Here's What Happened

Description

One agent doing everything sounds fine — until it doesn't. In this video we take the same Tokyo population task from Videos 3, 4, and 5 and split it across two specialised agents: a Researcher who can only search and store, and a Calculator who can only retrieve and compute. Neither can do the other's job. The role boundary is enforced at the tool level — not just in the prompt. This is CrewAI. And it's the last piece of the Orchestration layer in this LLMOps series. 📌 What's covered: → Why one generalist agent breaks down in production → CrewAI's three primitives: Agent, Task, Crew → Role design — how backstory and goal actually shape behaviour → Tool scoping — giving each agent only what their role requires → The context parameter — one line that handles the entire agent handoff → Process.sequential vs Process.hierarchical — when to use each → Reading the live output — clean execution vs Video 5's recovery behaviour → The built-in CrewAI trace link — free observability out of the box Full LLMOps Series — Orchestration Layer: Video 3 → Raw ReAct Agent (pure Python, 200 lines) Video 4 → LangChain AgentExecutor (66 lines) Video 5 → LangGraph (you own the graph) Video 6 → CrewAI Multi-Agent ← you are here #CrewAI #AIAgents #LLMOps #MultiAgent #Python #LangChain #LangGraph #GenerativeAI #MachineLearning #SoftwareEngineering