DEV Community

Omar Baruzzo
Omar Baruzzo

Posted on Originally published at omarbaruzzo.it

Most multi-agent systems are one problem, glued back together at runtime

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)