Breaking down like human teams
Multi-agent is powerful, but it also breaks down in ways very much like human teams.
TYPE 1
A game of telephone
Type one: a game of telephone. The request passes through three agents, and each summary loses a little. By the end, the result has drifted from the original idea.
TYPE 2
Errors add up
Type two: errors add up, called error compounding. Each step is ninety-five percent right, which sounds great. But chain ten steps together, and the whole thing is only about sixty percent right.
TYPE 3
Expensive and slow
Type three: expensive and slow. Fifteen times the tokens, for a job one agent could have done.
THE QUESTION
Can it be split?
So before using multi-agent, ask one question: can this job be split into independent parts? If yes, use multi-agent. If not, one good agent with enough tools is enough.
RECAP
The whole series in one picture
The whole series in short: AI uses tool calling to get things done. MCP is the shared socket. Install servers carefully. And multi-agent splits the work, when the job is big enough.
NEXT SERIES
Who watches the AI team?
Multi-agent is running for real now. Who keeps an eye on it? Watch the next series: Behind the scenes of an AI system.
This article is based on the video When multi-agent goes wrong from the Mark học AI channel. Watch the video (in Vietnamese) to see the animations.