3.6 — One agent or a team? The 3 paradigms

For a simple task, a single agent is enough. But for “build a complete website,” a single agent saturates. The modern solution: multi-agent orchestration — splitting the work among virtual specialists. Three main approaches (“frameworks”):

🟢 In plain words — a multi-tasking agent is like asking one person to be cook, waiter and cashier at once: fine for a tiny café, not for a big restaurant. Multi-agents means splitting the roles among several specialists who coordinate. The three frameworks below differ mainly in how they coordinate.

The “start-up” approach — CrewAI 🤝

The simplest and most intuitive. You create roles like in a company: a Manager, a Researcher, a Writer. The Manager receives the mission, instructs the Researcher, who passes notes to the Writer. (Strong enterprise adoption for its simplicity.)

The “strict factory” approach — LangGraph 🏭

The method for teams that want to avoid errors. Work moves along a graph (like an assembly line), each step strictly controlled. Key strength: a state memory (checkpointing) — if the system crashes, the agent resumes exactly where it stopped. It has become the production default.

The “group meeting” approach — AutoGen 💬

Agents are in a group chat: they debate to find the best solution. Very effective for code: a coder agent proposes, a tester agent replies “no, that doesn’t work, try again” until they agree.

🔔 Important update (late 2025): Microsoft merged AutoGen with Semantic Kernel into a new Microsoft Agent Framework (MAF), built for production (availability early 2026). AutoGen moves to maintenance (fixes only). The conversational paradigm described above remains valid and instructive; for a new project, you’ll now look at MAF.

🧭 The landscape moves fast. Other players exist (OpenAI Agents SDK, Google ADK, the agent-to-agent A2A protocol…). Don’t “marry” a framework: understand the 3 paradigms (roles / graph / conversation), the tools will follow.

🧩 Concrete example — “write a market report”: with CrewAI, a Researcher gathers the data and hands it to a Writer (roles, like in a company). With LangGraph, each step is a box in a graph, and if it crashes during writing, you resume there without redoing the research. With AutoGen, two agents debate the plan until they agree.

✅ In short

• A single agent often suffices; multi-agents are for tasks too broad for one.
• CrewAI = roles (most intuitive); LangGraph = graph + state memory (production default).
• AutoGen = conversation (strong on code); merged into MAF in late 2025.
• Remember the 3 paradigms (roles / graph / conversation), not a specific tool.

🔎 Going further (tech) — to code a lightweight agent yourself, the reference is smolagents (Hugging Face): its twist is that the agent writes Python code to call its tools (rather than JSON). The niveau-3-smolagents.ipynb notebook rebuilds it in a few lines, no API key.