Where the time actually goes in an AI coding-agent job

Andréas, a full-stack developer and CTO at a B2B SaaS, measured the runtime of AI coding-agent jobs on his agent-orchestration platform and found that infrastructure overhead costs 5–8 minutes per job, accounting for 20–30% of total runtime. By pre-warming worktrees, enabling persistent caches, and optimizing I/O, he reduced fixed overhead to under 60 seconds, leaving model inference as the dominant cost.

I run an agent-orchestration platform: it takes a ticket, spins up a git worktree, lets a Claude Code agent build the feature, and runs a deterministic verify gate before merging. A typical job takes 20–30 minutes. I used to blame slow model responses for the runtime, but stage-by-stage measurements proved that assumption wrong: infrastructure overhead costs 5–8 minutes of every job before the agent even finishes. Here is where the time actually goes and how to optimize it. On a standard dev machine Windows 11, Bun, Next.js , a single job pays a heavy infrastructure tax: | Stage | Measured Cost | Impact / Notes | |---|---|---| git worktree add | Seconds | Negligible overhead. | bun install | ~91s median | Cold worktree setup, even with a warm cache. | Lint & Typecheck | ~1–2 minutes | eslint + 2× tsc across the full repo with no shared cache. | next build | Minutes | The longest gate leg due to a cold .next directory every run. | Vitest | ~26 seconds | Fast and acceptable. | Model Time | ~15–22 minutes | Planning, execution, and review calls. | Fixed infrastructure costs account for 20% to 30% of total runtime . Eliminating this overhead reduces small job times from ~30 minutes to ~20 minutes, leaving the remaining time strictly bounded by model inference. Instead of creating a fresh worktree and running bun install from scratch for every job, maintain a small pool of pre-warmed directories containing: node modules directory invalidated only when the lockfile hash changes . .next , tsconfig.tsbuildinfo , .eslintcache .When a job starts, it claims a slot, checks out the target branch, and immediately executes. After completion, git clean resets the worktree while leaving cache directories intact. This eliminates the ~91-second install step on almost every job. Next.js supports persistent filesystem build caching. Combining this with a warm slot pool carries the .next cache between jobs, turning the longest gate step from minutes into tens of seconds for standard diffs. Adding --incremental to tsc and --cache to eslint within the persistent slot directory saves roughly 1 minute per job without changing verification outcomes. Real-time Defender scanning severely throttles small-file I/O operations like node modules and .next writes . Adding directory exclusions or moving worktrees to a ReFS Dev Drive improves install and build I/O speeds by 2× to 5× with zero code changes. Switching to bun install --linker isolated provides pnpm-style symlinked stores, yielding faster warm installs. However, caution is required: symlinks spanning worktrees can cause catastrophic unintended deletions if cleanup commands traverse outside the worktree root. tsc --incremental and eslint --cache .By optimizing the infrastructure pipeline, fixed job overhead drops from 5–8 minutes to under 60 seconds . The remaining runtime represents actual model reasoning—the exact phase worth waiting for. I'm Andréas — full-stack dev, CTO at a B2B SaaS, building my own agent tooling. Portfolio: https://andreas-bodin.vercel.app