Skill-based sub-agent routing cuts Claude Code API-equivalent cost 28% with zero quality loss
NoScene7932 · reddit · 2026-10-17
A developer built an agent-orchestration skill with Claude Code: Opus orchestrates and delegates subtasks to Haiku/Sonnet/Opus sub-agents based on task nature, cutting cost.
- Routing logic: the skill checks verifiability, judgment required, and wrong-answer risk. In tests, Opus assigned inventory functions and config conversion to Haiku, a date-parsing bug fix to Sonnet, and kept a money-allocation function with tricky edge cases for itself.
- Results: 15 runs with the skill vs 15 all-Opus runs showed 28% lower API-equivalent cost including the Opus orchestrator (mean cost per run $1.011 → $0.725); all 30 runs passed all 35 hidden checks. One four-task test; subscription allowance wasn't measured.
- The skill file and model-selection rules were written by Claude Code itself, then revised after running the comparison eval.
The skill is free to try in Claude Code, with full methodology and results published.
Related event: Using Opus as Orchestrator Cuts Claude Code Usage by 28%(3 posts)→
More from coding & agent
- Meta paper introduces "agent plasticity": best-performing agent isn't the most efficient learner — rohanpaul_ai · 2026-10-18
- Dev ships a bushfire prediction platform in a weekend with frontier models — CtrlAltDwayne · 2026-10-18
- Developer launches Is it vibe coded?, a site that detects AI-built websites by design fingerprints — sahilypatel · 2026-10-18
- 10 open-source GitHub repos that let AI agents control your phone, from Open-AutoGLM to mobile-use — victor_explore · 2026-10-18
- Indie dev's 7 AI projects in a year: hiring agent flopped, WeChat automation custom gig made ~$14K — sujingshen · 2026-10-18
- Dev with 108K followers: Codex is now slower than typing code by hand — lxfater · 2026-10-18