One good agent beats a team of agents

2025 sold an appealing idea: if one AI agent is useful, ten agents working together should be ten times better. In 2026, the teams running them in production landed somewhere less flashy. A crowd of agents doesn’t automatically outperform one well-designed agent.

Why multi-agent systems break

A Berkeley research group analyzed hundreds of multi-agent runs and published a failure taxonomy known as MAST (arXiv 2503.13657). Its fourteen failure modes roll up into three root causes — and none of them is “the model didn’t know”:

1. Vague specs. Each agent’s role isn’t defined tightly enough, so two agents do the same thing or nobody does what was missing.
2. Broken handoffs. Context gets lost when work passes from one agent to the next.
3. No verification. The result is accepted without anyone checking it.

The third one is the most expensive because it’s silent: one agent hands incomplete context to another, the second answers confidently and wrong, and not a single error shows up in any log. The customer just gets a bad answer.

The pattern that won: a coordinator, not a swarm

Of all the ways to organize multiple agents, the supervisor pattern became the production default in 2026: a coordinator routes each task to the right agent and reviews what comes back. The swarm — agents passing work directly to each other — feels nimbler, but on tasks that span several domains it burns more calls and more tokens to reach the same result.

The rule of thumb from teams in production: a second agent earns its place when it separates a knowledge domain or a distinct responsibility — not when it splits steps of the same task.

What that looks like in a synthetic agent

A well-designed synthetic agent looks more like a good role than a generic bot:

• Clear scope. What it resolves, what it escalates and what it never does without a person.
• Everything on the record. Every action is logged and reviewable.
• Measured on consistency. Getting it right once isn’t enough — it has to solve the same task the same way, again and again (τ-bench, arXiv 2406.12045).
• A supervisor. Another system, or a person, reviews what matters before it reaches the customer.

The bottom line

The useful question isn’t how many agents fit in your operation. It’s which role to cover first, how you’ll know it’s doing the job well, and who’s accountable when something goes wrong. Answer those three and one agent will outperform an improvised team.

Want to know which role to cover first? Talk to Luz.