Anthropic: A Single Agent Beats a Team of Subagents

Anthropic, the company behind the Claude models, has published new guidance on building AI agents. Subagents are separate AI helpers, each handling one specific task under the control of a coordinator. Anthropic now recommends folding subagents into a single agent loop instead. The company says modern models can handle far more complexity on their own. One client that switched to this approach saw roughly a 2x speedup. Claude Code is Anthropic's AI-based coding tool. Anthropic cut the system prompt for Claude Code, the baseline instructions given to the model, by more than 80% for newer models. The company also released a migration toolkit on GitHub, the code-hosting platform, so others could test the change themselves.

This calls into question the trend of building a separate subagent for every task or team. According to Anthropic, a single agent loop with skills that load only when needed works more reliably than a crowd of separate subagents. Anthropic notes that 30 to 50 tools in one prompt is a rough threshold. It is worth testing in practice, not guessing at. Teams can use the GitHub toolkit to check whether it makes sense to simplify their own agents.