Anthropic: 'use fewer subagents' — a single loop ran ~2x faster, and Claude Code's prompt was cut over 80%
First seen on X 22 hours ago@Xudong07452910 ♥ 1,8951 views
Anthropic is advising builders to fold subagents back into one main agent loop, arguing that complex multi-agent designs often hurt quality, cost and latency. In an Oct 22 post, it said a single agent with skills frequently beat both the one-prompt design and the subagent design across enterprise deployments; one customer saw runtime roughly 2x faster after switching to a single loop.
The reason, it said, is that shared context yields better answers — handoffs between agents become a lossy "telephone game." Anthropic said models from Opus 4.5 on can handle far more in one loop, and that it cut Claude Code's system prompt by over 80% for more advanced models. Subagents still fit verification, parallel work and permissioning, it added.
“Anthropic has recently started advising companies to "use fewer Subagents," even merging some expert Agents back into the main Agent Loop.”
Reactions (1)
- @peter_qiancc♥8
In most scenarios, a single agent can already solve the majority of problems. Most tasks don't reach the threshold where multiple agents become necessary.
