Stop telling Claude Code to use sub-agents: 288 runs, same recall, up to 221% more cost (Claude Codeに「サブエージェントを使って」と書くのはやめよう)
A · Strong evidence A pre-registered 288-run experiment by the Qiita author yunany90, with questions, hashed answers and scoring criteria registered before the first run and all run data published under CC BY 4.0.
Summary based on the full text·Original: about 4 minutes to read
In short
A Japanese Qiita post (6 November 2026) testing whether telling Claude Code to delegate search to its Explore sub-agent, at low, medium or high sub-agent effort, finds more. It does not. Across six open-source repositories and 12 "find every place" questions, recall stayed at 97-100% in all conditions while cost rose up to 221% and time doubled or tripled. Sub-agents always ran on the same model as the main agent. A side finding matters for scripts: in headless mode, the first result event usually lacks the answer.
Key numbers
- 288runs: 12 find-all questions x 2 models x 4 conditions x 3 repeats
- 97-100%share of the 151 target items found in every condition; differences under 2 points
- +74-116% / +143-221%cost per run increase when told to use sub-agents, Haiku 5.5 / Opus 5.5
- 2-3xrun time with sub-agents
- 210 of 216sub-agent runs whose first result event in headless mode contained no answer
As reported by the source.
Japanese·Web page·opens qiita.com
What it reports
- Repositories: undici, fastify, cachetools, Python-Markdown, anyio, zod; 151 target items extracted in advance by regex and parsers.
- No instruction: Haiku 98% at $0.0056 per run, Opus 100% at $0.16. With sub-agents: Haiku $0.0098-0.012 (+74% to +116%), Opus $0.39-0.51 (+143% to +221%), recall 97-99%.
- Raising sub-agent effort did not improve search; for Opus, medium cost more than high.
- Without an instruction, Claude Code never called a sub-agent on its own: 0 of 72 runs here, 0 of 504 in an earlier test.
- Sub-agents used the main agent's model: an Opus session spawns Opus sub-agents.
- In Claude Code 2.1.318 the Explore agent runs in the background, so
claude -pemits several result events. 210 of 216 sub-agent runs had no answer in the first one; a script reading only the first gets 0-9% of answers. Four Haiku runs did not exit for 20 minutes after answering. - The author first scored only the last result, noticed the issue, switched to concatenating all result events and logged the change in the pre-registration; both scorings are published.
How they got it
Claude Code 2.1.318 run headless with read and search tools only. Questions, answers (sha256) and a 15-point rubric were pre-registered on GitHub. Full data (scores, cost, time, sub-agent calls and models, answer text) are public at tokensave.app and GitHub.
Limits
- Little room for gains: both models already found nearly everything without help.
- Medium-sized repositories; large monorepos may differ.
- One instruction wording, one version, three runs per condition.
- The data page sits on a site the author appears to run (tokensave.app); no product claim is made in the post.
Why it matters for companies
Teams writing agent instructions often add "use sub-agents" as a default. For ordinary code search this test shows it only adds cost and time, and that the sub-agent inherits the expensive model. Leave effort settings at default unless your own measurement shows a gain. Anyone running Claude Code from CI or scripts should read all result events, or will silently score empty answers.
Key passages in translation
Japanese original
288回測りました。答えは変わらず、コストだけが増えました。
I measured 288 runs. The answers did not change; only the cost went up.
Japanese original
指示がないとき、Claude Codeが自分からサブエージェントを呼んだことは一度もありませんでした。
Without an instruction, Claude Code never once called a sub-agent on its own.
Japanese original
CIやスクリプトからClaude Codeにサブエージェントを使わせるなら、resultイベントはすべて読んでください。
If you make Claude Code use sub-agents from CI or scripts, read every result event.
Summary based on the full text. Read the original for the full argument and data.Read the original on qiita.com
- Tags
- Cost of AI