Anthropic urges single-agent designs over sub-agents

Anthropic recommended that companies check whether they could replace several specialist sub-agents with one lead agent that loads context only when needed. It said newer models, beginning with Opus 4.5 and including Opus 5.5 and Fable 5.1, could manage complex plans directly and could reduce cost and waiting time.

The company said its advice came from enterprise deployments examined by its Applied AI team. In one example, it said a customer-facing system used a dozen specialist sub-agents, which made brand consistency harder to maintain and increased latency.

Anthropic said some multi-agent designs still had a place when a single agent's context window degraded, when tasks had to run in parallel, or when tool selection became unreliable. It also said that on newer models it had cut the Claude Code system prompt by more than 80 percent because skills and tools could now be loaded only when needed.

To help teams change architectures, Anthropic released a migration kit on GitHub that walked users through identifying existing patterns, comparing alternatives and checking the result. The company also said teams should keep skill libraries from growing too large and warned that tool selection tended to worsen beyond 30 to 50 tool definitions.

The article said Anthropic's view aligned with recent research. It cited a Vals study that found no quality improvement from larger agent teams and quoted OpenAI developer Eric Provencher as questioning the value of agent swarms.