More AI agents isn't always better, new Google and MIT study finds

Minecraft planning tasks told the opposite story. Every multi-agent setup hurt performance by 39 to 70 percent. The problem: each crafting action changes the inventory state that later actions depend on. These sequential dependencies don't split well across agents.

Whenever each step in a task alters the state required for subsequent steps, multi-agent systems tend to struggle. This is because important context can get lost or fragmented as information is passed between agents. In contrast, a single agent maintains a seamless understanding of the evolving situation, ensuring that no critical details are missed or compressed during the process.

Three factors that tank multi-agent performance

Tasks with many tools, like web search, file retrieval, or coding, suffer most from multi-agent overhead. The researchers say splitting the token budget leaves individual agents too little capacity for complex tool use.

Capability saturation matters too. Once a single agent hits about 45 percent success rate, adding agents brings diminishing or negative returns. Coordination costs eat up any gains, according to the researchers.

Error accumulation is the third problem. Without information sharing, errors compound up to 17 times faster than with a single agent. Mistakes pass through unchecked. A central coordinator helps; errors "only" increase by a factor of four, but the problem doesn't go away.

The 45 percent threshold

The key rule of thumb: if a single agent solves more than 45 percent of a task correctly, multi-agent systems usually aren't worth it. Multiple agents only help when tasks divide cleanly. For tasks needing around 16 different tools, single agents or decentralized setups work best.

Model providers showed slight differences. OpenAI did well with hybrid architectures, Anthropic with centralized ones. Google proved most consistent across all multi-agent setups.

The researchers also built a framework that correctly predicts the best coordination strategy for 87 percent of new configurations, what they call "a quantitatively predictive principle of agentic scaling based on measurable task properties."

Single agents use tokens more efficiently

The researchers tracked tasks completed per token budget. Single agents averaged 67 successful tasks per 1,000 tokens. Centralized multi-agent systems managed just 21; less than a third. Hybrid teams completed only 14 tasks per 1,000 tokens.

The culprit is coordination overhead. Hybrid systems need about six times more reasoning turns than single agents. The researchers recommend three to four agents maximum when budgets are tight.

For developers, the message is simple: start with a single agent. Only switch to multi-agent systems when the task splits into independent pieces and single-agent success stays below 45 percent.