Multi-agent is the default reach right now, and most of the time it's the wrong one. Splitting a task across agents doesn't add capability — it moves the hard part from inside the agents to the glue between them.
What the glue costs you:
- Every handoff drops or distorts context.
- Each agent owns its control flow, so nondeterminism multiplies instead of adding.
- Errors compound: five agents in a chain pass mistakes downstream as facts.
- No stack trace for "they misunderstood each other".
- More prompts to maintain, more surfaces to evaluate, more cost per run.
When it's actually justified: genuinely independent, parallelizable subtasks, or parts that need different tools, permissions, or trust boundaries. That's a design reason. "More agentic" isn't.
Default to one capable agent with good tools and a fixed workflow around it. It's cheaper, testable, and you can reason about what it did.
Full version: https://www.omarbaruzzo.it/en/blog/meno-agenti-di-quanti-pensi
Top comments (0)