Anthropic guide: fold unnecessary subagents into skills and tools, with 2x faster execution reported
omarsar0 · x · 2026-10-25
omarsar0 reviews Anthropic's new guide "Reducing subagents over time": as frontier models get better at context management, Anthropic now recommends folding unnecessary subagents back into the main agent loop, with one customer seeing roughly 2x faster execution.
The guide outlines three paths (see diagram):
- Merge subagents: combine those sharing context into fewer ones
- Call them as tools: main agent invokes subagents via tool calls, keeping the conversation
- Fold into skills and tools: prompts become skills, API calls become tools
Rules of thumb: Skills teach the agent HOW (procedures, expertise); Tools/MCP let it DO things (APIs, live data); Subagents only for parallel work, isolated context, or tasks a single agent struggles with.
The author still uses subagents for complex research (parallelism, own context windows, dynamic workflows + /goal), but agrees newer models' persistence and large-context handling make subagents often get in the way. Key caveat: don't rip them out blindly—run evals on cost and efficiency to decide whether and how to fold them.
More from coding & agent
- print3d: open-source agent skill turns photos or text into Bambu Lab 3D prints — udmrzn · 2026-10-25
- Dev launches gaming-focused remote MCP server for squad-finding, seeking testers — Rashad8705 · 2026-10-25
- Redditor compares the agent boom: Claude Code, OpenClaw, Hermes vs hosted rivals — Odd-Aside456 · 2026-10-25
- Jev ecosystem explodes: 5 open-source apps from a 7.1s flight-search agent to SQL-in-English — udmrzn · 2026-10-25
- PhD student says Claude now does most of his thesis work, down to the draft — GarrisonLovely · 2026-10-25
- Google's Jon Barron shares his full AI paper-writing workflow with Gemini — lmoroney · 2026-10-25