Multi-agent
Only worth it when a single agent can’t handle the job, because coordination has a cost.
In one sentence
Architecture with multiple specialized agents cooperating: one researches, one writes, one reviews. It delivers more power, but also more coordination work, so it’s only worth it when a single agent can’t handle the job.
Before
One agent can’t handle it
What it does
Specialized agents cooperating
After
More power and more coordination
Example
Three agents: one reads the report and extracts facts, one writes the letter, one checks the letter against the extraction. It works better than a single agent, but it costs three calls, three failure points, and a protocol for who talks to whom.
For 4,000 letters a year, the gain pays off. For 40, it wouldn’t.
The common mistake
Starting with multi-agent because it seems more advanced. The rule is the opposite: use one agent until it hurts, and only then split it up. Each new agent adds coordination, which is where agentic systems usually break.
In practice
- Only split when a single agent has already failed in a measured way. Intuition isn’t enough.
- Each agent should have input and output in a fixed schema, like a contract between them.
- Measure the cost of coordination: extra calls, extra latency, extra points of failure.
How to make it tangible
Two or more agents with defined inputs and outputs between them, and a document saying who calls whom.
Connects with
Translated from Portuguese with AI assistance.