One AI agent often outperforms a team, Anthropic says
In brief
According to Journal du Net, Anthropic said on 22 October that one agent with skills often beats multi-agent setups on quality, cost and speed. In June 2025 it had described a research tool with several agents that scored 90 percent better than one agent on its tests, at much higher compute cost. A client saw the system run about twice as fast after merging everything into one agent.
Source : Journal du Net - « Un seul agent vaut mieux que plusieurs : ce qui a changé, et pourquoi »
What it changes for an SME
If you plan AI agents for follow-up, support or monitoring, giving each team its own agent is no longer what Anthropic recommends as the default. One agent that loads skills only when needed can cut compute cost, delays and information loss at each handoff. Separate agents still make sense when memory is overloaded, when several jobs must run at once, or when there are too many tools to choose from. Human oversight remains a risk: according to Anthropic's IPO prospectus dated 26 October, the speed and volume of agent actions can tire reviewers, leave them without context, or delay checks.
What you can do
- List current or planned AI agents and check whether they simply mirror your departments rather than a real need.
- Try one agent with a small set of skills before adding specialized agents.
- Decide who reviews outputs and how often, so volume does not outrun your capacity to check.
Let's talk about the time you could save.
In 20 minutes, we find where AI could save you time or money. Free, no commitment.