Anthropic: Single agents with skills often beat subagent designs in enterprise deployments
addyosmani · x · 2026-10-27
Addy Osmani highlights Anthropic's new best-practices writeup on when to fold subagents into the main agent loop.
Key finding: across several enterprise deployments, Anthropic's Applied AI team found a single agent equipped with skills frequently outperformed both the one-big-prompt design and multi-subagent orchestration on quality, often at lower cost and latency. One company's customer-facing agent had a dozen specialist subagents that made brand consistency hard to maintain and added latency.
Osmani's rule of thumb: before adding a subagent, ask what problem it solves—a bloated system prompt is often the real issue, and skills plus tool search can fix that without splitting context. Keep a subagent only when work is parallelizable, needs fresh context, or the main agent genuinely can't pick the right tools. Otherwise fold it in, turn its prompt into a skill, run evals, and keep what wins.
Subagents still shine for verification, parallel work, and modular permissioning. Anthropic ships a migration kit with a guided workflow you can rehearse on a mock system first.
More from coding & agent
- Writing a 994k-subscriber newsletter with Claude Code: cutting 8 of 10 words to train 74 rules — rubenhassid · 2026-10-27
- shadcn: a voice dictation app is just T3 code plus one prompt — shadcn · 2026-10-27
- Coding got cheap, but the friction we removed was doing useful cognitive work — charles_irl · 2026-10-27
- One dev + AI builds open-source Rust suite targeting Photoshop, Premiere and Office — ankur_maker · 2026-10-27
- Claude Code's idle compaction annoys users — here's the setting to turn it off — JasonBotterill · 2026-10-27
- Creator builds CCTV-style video intro with Claude Code, still stuck on AI dubbing — xiaohu · 2026-10-27