The 'Short Leash' Method: Expert Devs Beat Frontier AI by Never Looking Away

A maintainer of security-critical software at okTurtles lays out a workflow for using AI coding agents without surrendering code quality or your own understanding of the codebase. The core argument is that the fashionable ‘vibe engineering’ setups — fleets of parallel agents run by an orchestrator while the developer relaxes — produce slop and make it impossible to build genuine understanding of what’s being shipped. Frontier models like Fable 5 still generate inefficient, ugly code, especially in niche domains with sparse training data, because they don’t reason beyond what they’ve seen.

The alternative, the ‘short leash’ method, is aimed squarely at expert developers who already outclass the models in their domain. It means no YOLO/skip-permissions mode, an agent that surfaces a diff on every proposed change, and a human who actually reads each diff, denies bad edits, and intervenes the moment the agent drifts off course. Planning and task-tracking come first; commits happen after every subtask so the agent can’t silently destroy prior work. Throughout, the permission prompts double as a mechanism to keep the developer’s mental model of the code current.

On review, the piece argues both a human and an AI should examine every PR — the AI acts as a fast linter for common mistakes while the human catches higher-level and directional issues. It also pushes process norms: PRs should carry an ‘AI Disclosure’ naming the exact models used, and the human author must review AI-written code line-by-line before requesting maintainer attention, on the principle that an AI-assisted PR is really the AI’s work that the human is accountable for understanding.