One AI, twenty competitors
Give one AI a big job: research twenty competitors, reading a dozen web pages for each. By competitor number ten, it starts forgetting competitor number one.
WHY
The desk has a limit
Because each AI has a limited desk, called the context window, as video 9 explained. Pile on too many documents, and the desk overflows.
SPLITTING THE WORK
Orchestrator and subagents
The fix: split the work among many agents, called multi-agent. A lead agent, called the orchestrator, makes a plan and gives each subagent a few competitors. Each subagent has its own context window, they run at the same time, and when they're done they only send back a short summary.
REAL NUMBERS
Anthropic tried it
Anthropic tried this for Claude's research feature. An orchestrator with several subagents did up to ninety percent better than a single agent, on their internal test.
THE COST
A big team is expensive
But a big team is expensive: in total it uses about fifteen times the tokens of a normal chat. So it's only worth it for big jobs that can be split into independent parts.
NEXT
Who's in charge?
Now there's a team. Who's in charge, and who checks the work? Part six.
This article is based on the video Why multi-agent from the Mark học AI channel. Watch the video (in Vietnamese) to see the animations.