Anthropic: most agent workflows should fold subagents into a single agent loop
_AustinCalvert_ · x · 2026-11-17
Anthropic's Applied AI team reports that across several enterprise deployments, a single agent with skills frequently beat both one-prompt-for-everything and multi-subagent designs on quality, at lower cost and latency. Key advice: don't model agents on your org chart; put plans in the main agent's context instead of plan mode; design MCP tools around jobs-to-be-done rather than mirroring APIs; evaluate on full task completion with coherent, non-conflicting evals; and if you keep subagents, call them as tools so the main agent owns the context. A migration kit is provided to guide folding routers, planners, and domain subagents back into the main loop.
More from coding & agent
- Amanda Launches API That Sends AI Agents to Meetings for You — Scobleizer · 2026-11-17
- Fuzzy UI: agents could let users answer roughly instead of exactly — _AustinCalvert_ · 2026-11-17
- Building 3D Worlds With Claude: The Case for a 'Signature Demo' You Keep Upgrading — Kyrannio · 2026-11-17
- GBrain's hosted memory layers personal, team, and company brains across ChatGPT, Claude, and Cursor — garrytan · 2026-11-17
- Build a Multi-Agent System From Scratch ships chapters 9-10 on subagents and A2A — _nerdai_ · 2026-11-17
- Break-even math for 20-turn agent loops: caching and routing cut API burn 76% — TokenMarginDev · 2026-11-17