Quality assurance · Production

Mercari's pj-double: Agent-Spec Driven Development achieves 150%+ development speed improvement

The problem

AI usage at Mercari was fragmented — individual developers used different approaches (Vibe Coding) that could not be reproduced or shared organizationally, widening the gap between AI-proficient and non-proficient developers and preventing company-wide productivity gains.

First attempt

Early standardization attempts were limited to sharing CLAUDE.md and Cursor Rules locally, without reaching organizational scale. A build trap in QA tooling — delivering a rich UI app instead of minimal prompt infrastructure — raised contribution barriers and slowed method iteration.

Workflow diagram · grounded in source
1
Developer defines task scope
trigger
“最初に目的・ステップ・完了条件さえ与えれば、Agentが自律的にタスクを完遂するため、開発者はその結果のみを評価すれば良い”
2
Agent Spec generation
ai_action
“メルカリのナレッジ基盤に最適化されたエージェントが自動的に一次情報にアクセスし、詳細な実装計画を生成します”
3
Standards and security validation
validation
“別のエージェントが、実装計画がサービスのコーディング規約に沿っているか、計画が所定のセキュリティ観点をクリアしているかなどのさまざまな調査を行います”
4
Developer clarification
human_review
“この2つのエージェントが交互に修正と評価を繰り返し、最後に要件や仕様における不確実要素が残った場合には、開発者に追加の質問を行います”
5
Autonomous implementation
ai_action
“実装を行うエージェントと、テストや静的解析を行うエージェント、実装結果のAgent Specとの整合性を検証するエージェントが互いに協調しながら、タスクの完了まで自律的に実行されます”
6
Weekly productivity tracking
feedback_loop
“週例のミーティングにおいて細かく開発フェーズごとに行なった作業と要した時間を報告してもらいました”
Reported outcome

SDD-based projects averaged over 150% development speed improvement versus estimates, and over 80% improvement versus other AI-assisted methods.
pj-double expanded from one person to a team of over ten and scaled company-wide across Mercari's full delivery cycle.

Reported metrics
development speed improvement (SDD vs original estimate)150%以上
development speed improvement (SDD vs other AI methods)80%以上
Development efficiency (best-case example)5分の1
Pj-double team size growth1人から10人超のチームに拡大
Reported stack
Claude CodeDX
Source
https://engineering.mercari.com/blog/entry/20251201-pj-double-towards-ai-native-development/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

SDD-based projects averaged over 150% development speed improvement versus estimates, and over 80% improvement versus other AI-assisted methods.

What tools did this team use?

Claude Code, DX.

What results were reported?

development speed improvement (SDD vs original estimate): 150%以上; development speed improvement (SDD vs other AI methods): 80%以上; Development efficiency (best-case example): 5分の1; Pj-double team size growth: 1人から10人超のチームに拡大 (source-reported, not independently verified).

What failed first in this deployment?

Early standardization attempts were limited to sharing CLAUDE.md and Cursor Rules locally, without reaching organizational scale.

How is this quality assurance AI workflow structured?

Developer defines task scope → Agent Spec generation → Standards and security validation → Developer clarification → Autonomous implementation → Weekly productivity tracking.