Coding Agents, Governed

What actually goes wrong

• Every engineer improvises. Custom prompts, custom permissions, no shared definition of done — results are unreviewable and unrepeatable.
• Agents run with too much access. Full credentials, production reach, and a prompt that says "be careful".
• No gate. Agent-written code merges like human code, except nobody fully read it.
• No measurement. Leadership cannot answer whether the agents help, because nothing was ever counted.
• Un-isolated environments. Agents experiment where they deploy. When the workspace is production-adjacent, "it works on my machine" becomes an outage.

If I were doing my agentic development in a non-isolated environment, I would be cooked by now.— Jan Lafko, independent developer, on the Turbo Flow workflow

The governed shape

• Isolation. Agent work happens in worktrees and disposable environments — never directly on shared state.
• One workflow. Plan → research → PRD → ADRs → build, the same for every agent and every human.
• A review gate. Cross-model review on every PR, then a human merge decision. Details.
• A constitution. The rules agents operate under are written, versioned, and enforced — see agent governance.
• Metrics. PRs, review rejections, tokens, cost, cache hit rate — published, not guessed.

What the metrics say

My own rig runs this workflow against every repository I own. For the audited week of September 21–28, 2026: 180 merged PRs across 16 repositories, 967 gate reviews, 6.57 billion tokens, 96.5% prompt-cache hit rate. The agents build; the gate catches; a human merges. Nobody gets paged.

Figures below come from one audited production week (September 21–28, 2026) across my own GitHub account — merged-PR counts, gate logs, and token accounting, published weekly on the Turbo Rig stats page. They are measurements of my own workflow, not industry averages.

The narrative version with receipts: Turbo Rig case study. The environment that makes it ergonomic: Turbo Flow.