Anthropic now tells companies to drop unnecessary subagents; single agent with skills 2x faster
_AustinCalvert_ · x · 2026-10-25
Anthropic is now recommending companies remove unnecessary subagents — a notable shift in agent architecture. Its team found a single agent equipped with skills frequently outperforms complex multi-agent systems, with one customer seeing 2x faster execution. The emerging rules of thumb: Skills teach the agent how to do things; Tools/MCP give it the ability to act via APIs and live data; subagents only for parallel work, isolated context, or tasks a single agent struggles with. Progressive disclosure is the interesting mechanism. Max Rozen's viral AGENTS.md line "DO NOT USE SUBAGENTS EVER" (557K views) now looks prescient.
More from coding & agent
- AI productivity booms may owe as much to workflow changes as to AI itself — _AustinCalvert_ · 2026-10-25
- From Microservices to Agents: InfoQ Maps the Next Evolution of Distributed Systems — bibryam · 2026-10-25
- Running 30+ GrokBot agents as a team: loops, graphs and harness engineering — 2C_ornot2C · 2026-10-25
- Vicem: a zero-dependency local MCP server for code skeletonization and agent handoff — NoConclusion8361 · 2026-10-25
- Code is clay now: an iOS developer on how agentic coding reshapes software building — bibryam · 2026-10-25
- Why agent SKILLs are worth the tokens: a practical breakdown — ievkz · 2026-10-25