Anthropic says one AI helper often beats a team of many

AI company Anthropic is telling businesses to check whether they have too many AI helpers. Its advice is to see if a team of specialized helpers can be replaced by one main helper that pulls in information only when it needs it. In one comparison, the single-helper setup ran about twice as fast. Anthropic also released a free migration kit on GitHub to help teams test the change.

What changed

In the AI world, a helper that carries out tasks is called an agent. Many companies built one main agent that hands work to smaller specialist agents, called sub-agents. Anthropic says newer models, starting with Opus 4.5 and including Opus 5.5 and Fable 5.1, can handle complex plans on their own in a single loop. That can also lower costs and shorten wait times.

The advice comes from Anthropic's Applied AI team, which worked on several enterprise deployments. It found that one agent using "skills" often beat both one giant instruction block and setups with many sub-agents. Skills are packaged workflows that load only when needed.

How teams end up with too many helpers

One company had a dozen specialized sub-agents inside its customer-facing agent. That made it hard to keep a consistent brand experience, and replies were slow. Anthropic compares this to Conway's Law: the AI setup ends up copying the company's org chart, with each team shipping its own helper. The result can be hundreds of overlapping agents with no shared standards, plus a "Frankenprompt," where teams keep adding rules that contradict each other.

When many helpers still make sense

Anthropic does not say sub-agents are obsolete. It names three cases where they still fit:

• A single agent shows measurable performance decline as its working memory fills up.
• Several tasks need to run at the same time in separate working spaces.
• A single agent struggles to pick the right tool for the job.

The risk of merging too fast

Anthropic warns against merging blindly. In one test with a support agent, folding in a sub-agent did cut wait times, but it broke a process that had been working. The main agent began answering on its own instead of consulting the right specialist. The migration kit walks teams through finding their current patterns, comparing alternatives and checking which approach works best.

What to do

Anthropic suggests a small central team own the main instructions and overall quality. It also says tool choice tends to get worse beyond 30 to 50 tool definitions, so each team should have an owner keeping its skill library from growing too large. Reference material, such as guidelines or product specs, should sit in a knowledge base the agent searches only when needed.

Outside research points the same way. A study by Vals found agent teams can be faster than single agents but do not deliver higher quality. Going from ten to 30 or 100 agents brought no quality gains and much higher costs. OpenAI developer Eric Provencher called agent swarms "almost certainly" wasted money.