Audit a 12-layer LLM agent stack for regressions, memory pollution, tool misuse, and hidden loops.

Design agent action spaces, tool definitions, and observation formats to improve task completion rates.

Diagnose AI agent failures with structured self-debugging, containment, recovery, and introspection reports.

Trim ECC to a repo's needed skills by sorting them into daily and library buckets.

Operate as an agentic engineer with eval-first execution, task decomposition, and cost-aware model routing.

Build a persistent multi-agent operating system in Claude Code with specialists, commands, memory, and scheduling.

AIエージェントが大量の実装を生成するチームの開発とレビューの運用を設計する。

ジョブやフック、MCPサーバーの稼働状況と冗長性を証拠に基づいて監査する。

Claude Codeに永続メモリやスケジュール実行を加え、常時稼働の自動エージェントを構築する。

単純なパイプラインからマルチエージェントDAGまで、自動Claude Codeループを設計する。

複数セッションのプロジェクトを、自己完結したステップ付きの構築計画に変換する。

Persist project memory across Claude Code sessions, auto-loading context and tracking git activity deterministically.

Orchestrate parallel coding agents in isolated worktrees, monitoring their progress and reading structured reports.

Audit Claude Code context window usage and get prioritized recommendations for cutting token waste.

Extract reusable skills from sessions via a legacy Stop hook; deprecated in favor of v2.

Build atomic instincts from session hooks, scoped per project to avoid cross-project contamination.

Run a complete verification system for Claude Code sessions to check work before finishing.