Infrastructure overhead, not AI models, eats 20–30% of coding-agent job time

A developer running an AI coding-agent platform discovered that infrastructure setup — not model inference — consumes 5–8 minutes of every 20–30 minute job. Measurements showed that dependency installation, linting, type-checking, and cold Next.js builds collectively account for the bulk of fixed overhead before the Claude Code agent even completes its work. To address this, the developer proposes maintaining a pool of pre-warmed worktree directories with persistent caches and pre-installed dependencies, avoiding redundant setup on each job. Additional gains come from enabling incremental TypeScript and ESLint caching, persisting Next.js build output across runs, and excluding worktree directories from real-time antivirus scanning. Applying these optimizations is projected to cut fixed overhead from 5–8 minutes to under 60 seconds, reducing total small-job runtime from roughly 30 minutes to around 20 minutes.