Anthropic says cut most subagents: single-agent loop ran ~2x faster in enterprise tests

Arindam_1729 · x · 2026-10-27

Anthropic's applied AI team compared org-chart-style multi-agent setups (router + planner + specialist subagents) against single-agent designs across enterprise deployments. The counterintuitive finding: a single agent with skills frequently won on quality, cost, and latency — one customer ran 2x faster after switching to a single agent loop.

The culprit is handoffs: every subagent returns a summary, and every summary drops detail — Anthropic calls it a telephone game. Their recommended loop instead:

The same author also open-sourced OpenDuet (MIT), an open-models version of Claude Duet: proactive morning briefings, app monitoring while you're away, pings only when needed; anything that sends, posts, or pays waits for approval. Each bot runs its own open model with an editable MEMORY.md, built with Mastra, Nebius and Composio.

Related event: Anthropic's New Guidance: Fold Subagents Into Your Main Agent Loop(11 posts)→

Original post →

More from coding & agent

coding & agent channel →