Like a newsroom
A newsroom has an editor-in-chief, reporters and proofreaders. Multi-agent splits roles the same way.
ORCHESTRATOR
Give clear assignments
The orchestrator takes the job from you, breaks it down, and writes clear instructions for each subagent: what to find, where, and in what format. Vague instructions mean subagents end up doing the same work twice.
SUBAGENT
Each worker has its own toolkit
The subagents are specialists. Each has exactly the tools it needs: one searches the web, one reads spreadsheets, one writes.
REVIEWER
Another pair of eyes
Add a critic agent, called a reviewer: it writes nothing, it just looks for mistakes, checks sources and points out gaps. Two different pairs of eyes catch errors one would miss.
PUTTING IT TOGETHER
One result for you
Finally, the orchestrator puts everything together into one result for you.
A SHARED LANGUAGE
MCP and A2A
Agents talking to each other need shared standards too. MCP lets agents plug into tools. And A2A, short for Agent to Agent, is the standard for one company's agent to hand work to another company's agent.
NEXT
Teams get messy too
But every team gets messy sometimes. The last part: when multi-agent goes wrong.
This article is based on the video How multi-agent teams work from the Mark học AI channel. Watch the video (in Vietnamese) to see the animations.