When it is not one AI agent doing the work but several at once. Sometimes they complement each other, sometimes they get in each other's way.
You send five people to buy a present for the same birthday without letting them talk. They come back with five scarves. None of them got it wrong alone. The organisation did.
In June 2025 Anthropic described how it built its research system. The design is simple. A lead agent takes the question, splits it and sends smaller agents to search in parallel, each in its own direction. Then the lead puts it together. According to the company, on its internal research test this setup beat a single Claude Opus 4 agent by 90.2%.
The pleasure is not cheap. Again according to Anthropic, a single agent uses about four times more tokens than a normal chat, and a multiagent system about 15 times more. Tokens are the pieces of text the model's work is billed by. More agents, bigger bill.
And the errors look exactly like the present story. Early agents spawned 50 subagents for simple questions, searched endlessly for sources that do not exist and distracted each other with too many updates. Without a detailed task description they duplicate work and leave gaps. That comes from Anthropic itself.
Outside researchers see the same thing. In a 2025 paper, a team described 14 ways these systems fail and tracked them across more than 1,600 traces from seven popular frameworks. They grouped them in three: badly designed systems, misalignment between agents, and weak checking of the final result.
When it is worth it
Anthropic itself says where it does not fit: when all the agents need to see the same thing or depend heavily on each other. Most coding tasks, in its words, have fewer truly parallel pieces than research does. Before you launch a swarm, ask whether the task really splits into pieces. If it does not, one agent is cheaper and easier to watch.