Anthropic says drop unnecessary subagents: single agent with skills runs 2x faster
_AustinCalvert_ · x · 2026-10-25
Anthropic now recommends companies REMOVE unnecessary subagents from their AI systems — a significant agent-architecture shift. Its team found a single agent equipped with skills frequently outperforms complex multi-agent setups, with one customer seeing roughly 2x faster execution.
The emerging rules of thumb:
- Skills: teach the agent HOW to do something — procedures, expertise, instructions.
- Tools / MCP: give the agent the ability to DO things — APIs, actions, live data.
- Subagents: reserve for parallel work, isolated context, or tasks a single agent struggles with.
The key mechanism is progressive disclosure. Varunentropy adds context: in early 2026 Claude Code, Codex and OpenCode all shipped Explore/context-gathering subagents to fix context rot, but the default has shifted to a main agentic loop + progressive disclosure — harness design must keep evolving with frontier model behavior and stay lean.
More from coding & agent
- Apple & CMU's SSR cuts multimodal agent reasoning latency by 90% — jiqizhixin · 2026-10-25
- Graphify, a 135K-star repo, maps codebases into queryable knowledge graphs for AI agents — techNmak · 2026-10-25
- Matt Pocock: deterministic checks could safely automerge refactor PRs under strict rules — mattpocockuk · 2026-10-25
- Streamrunner ships: AI-powered native macOS reimagining of OBS Studio for $39.99 — jasonkneen · 2026-10-25
- Dev shares crude cross-interface agent setup: Linux box with limited apps plus CUA monitoring — tokenbender · 2026-10-25
- Dev uses agents to monitor group chats and send greetings on his behalf, eyeing voice cloning next — tokenbender · 2026-10-25