Hacker News front pageJul 3, 2026, 03:11 UTC+8

The Short Leash Method: Using AI agents for security-critical code without sacrificing quality

The Short Leash AI Coding Method for Beating Fable

Greg Slepak from okTurtles outlines a method where the developer stays in the loop at all times—reviewing every diff, denying permissions when the agent veers off, and committing after each subtask. He rejects fully autonomous 'vibe engineering' and argues that even non-frontier models can beat Fable 5 this way. For reviews, AI acts as a fast linter while the human catches directional issues; the PR author must do a line-by-line self-review and disclose which model was used.

Why it matters: A developer experience post with a concrete method and a clear stance, not marketing fluff. Hits all three HKR axes, but the author and platform aren't industry headliners, and the topic is dev-tool workflow rather than an industry-level event. Policy places it right at the fe...

Read the original ↗Export Markdown

Take

Greg Slepak spent over a year figuring out how to use AI agents without letting them wreck the codebase. His answer is the Short Leash method: review every diff before it lands, deny permissions the moment the agent veers off, and commit after each subtask so a rogue model can't delete earlier work. It's the polar opposite of the YouTube-famous vibe engineering where a dozen agents run in parallel while you sip coffee.

He makes a concrete claim: Fable 5 produces working but ugly, inefficient code, especially in niche domains with sparse training data. His experience says a developer who stays in the loop can beat that quality even with non-frontier models. The review section is equally practical—AI acts as a fast linter catching common mistakes, the human catches directional issues, and the PR author must do a line-by-line self-review and disclose which model was used.

I'd discount this a bit because the post doesn't provide benchmarks or side-by-side comparisons; "beating Fable" is personal judgment, not measured. But if you're already shipping production code with agents, this discipline is more actionable than any new model release. The useful bit isn't the tooling—it's the commitment to never letting the AI work while you look away.

Source: Hacker News front page · blog.okturtles.org
Published: Jul 3, 2026, 03:11 UTC+8 · Jul 3, 2026