10 Prompting Mistakes That Only Show Up in Multi-Agent Systems — And How to Fix Each One

Here is a failure mode I've watched happen to competent engineers more than once.

They've read the literature. They know how to prompt. They've built single-agent setups that work well — coherent outputs, consistent behavior, no obvious disasters. They take that same prompting instinct into a multi-agent pipeline and something goes wrong that shouldn't.

Not hallucination. Not a wrong answer. The agents each did something reasonable. The system produced garbage anyway.

The problem isn't the model. It's that prompting for a multi-agent system requires different thinking than prompting a conversation. In a chat, you're always present. You can clarify, redirect, catch drift in real time. In a multi-agent system, you write the instructions once, and every agent acts on them without you — in sequence, in parallel, often across many invocations before you see any output.

Every ambiguity compounds. Every missing constraint gets filled by the model's best guess. Every vague word in a system prompt is a decision your agent makes without you.

Here are the ten most common ways that goes wrong — and the concrete fix for each.