Engineering1 min read
Why multi-agent beat one big prompt
A single agent with twelve tools looks simpler until the first conversation crosses a domain boundary.
The first version of the sales bot was one agent holding every tool. It worked in demos and failed the moment a user asked about shipping halfway through building a cart.
What actually breaks
Tool selection degrades as the tool list grows. With twelve tools in one prompt the model picks plausibly rather than correctly, and there is no seam to test against.
Each agent owns three or four tools, not twelve.
A supervisor owns routing and shared state.
Every agent is independently testable.
The trade-off
You pay in latency — routing adds a hop — and in orchestration code you now own. Below roughly five tools, one agent is still the right answer.
graph = StateGraph(AgentState)
graph.add_node('supervisor', supervisor)
graph.add_node('cart', cart_agent)
graph.add_conditional_edges('supervisor', route)- LangGraph
- RAG