Anthropic tells companies to use one agent, not many helper agents

Anthropic makes Claude. It tells companies to check if they can replace many helper agents with one primary agent. This comes from its Applied AI team and its work with many companies. Anthropic tells that new models can do hard plans without helper agents, and this can decrease cost and wait time. The company gave no full test data.

How it works

The one agent uses skills, which are workflows that the agent reads only when a task uses them. It gets tools and guides only when a task uses them. Anthropic cut the Claude Code system prompt for new models by more than 80 percent. A small team must control the system prompt and the quality.

The numbers

Anthropic tells that one agent was two times as fast in one test. A study by Vals found no quality increase from more agents, and its cost was more. Anthropic tells that the model selects tools with less precision above 30 to 50 tool definitions.

What is not known

Anthropic tells teams not to merge without tests. In one test with a customer agent, a merge decreased wait time but stopped a procedure that operated before. Helper agents continue to help when:

• Quality falls in one context window.
• Tasks must operate in parallel.
• The agent selects tools with errors.

Insight: Anthropic pushes skills over agent counts

Anthropic tells teams to make one agent with skills. Anthropic also gave its Agent Skills standard to a neutral group. Novastart showed a system with a maximum of 10 helper agents. This reads as a move to make skills, not agent counts, the unit that companies use. An open standard helps Anthropic set it. Check which skill format your tools use.