Developers Propose ‘Short Leash’ Method to Prevent AI Coding Agents From Going Off Rails
A security-focused researcher has introduced the ‘Short Leash’ AI coding method to prevent autonomous agents from generating inefficient code or losing codebase context.
- The Short Leash method requires expert human developers to remain actively in the loop, rejecting fully autonomous ‘YOLO’ modes.
- Developers must manually analyze code diffs in permissions prompts rather than letting multi-agent systems run unattended.
- The methodology was developed by Greg Slepak of okTurtles after a year of testing custom forks of AI coding agents like Crush.
- According to the source, even advanced models like Fable 5 generate ugly, inefficient code when left to operate without strict human oversight.
Software developers can beat the performance of advanced models like Fable 5 by keeping AI agents on a strict, human-monitored protocol. According to Greg Slepak, a protocol developer and maintainer at okTurtles, the newly detailed “Short Leash” method rejects the popular trend of hands-off automation in favor of continuous, manual code review. By forcing agents to display exact code diffs and requiring explicit human permissions for every change, developers can prevent AI tools from writing inefficient code or drifting entirely off course during complex tasks.
How does the Short Leash AI coding method work?
The Short Leash methodology is designed specifically for expert software engineers working on security-critical systems who want to leverage AI without sacrificing code quality. The process begins with a structured planning phase to research the task and break it down into trackable steps. Crucially, the developer never enables “YOLO” mode, which dangerously skips permissions, and never allows the AI to write code unattended. Instead, the developer sits and actively analyzes the precise diffs proposed by the agent. This continuous feedback loop ensures the engineer’s understanding of the codebase remains fully up to date while catching errors before they are committed.
Why do fully autonomous AI agents fail in software development?
Many popular tutorials advocate for complex setups featuring up to 12 parallel agents managed by an orchestrator. Slepak criticizes this “vibe engineering” approach, calling it “slop writing and reviewing slop.” When developers remove themselves from the loop, they lose their own comprehension of the codebase. Furthermore, according to Slepak’s research, AI models cannot think beyond their training data. When tasked with niche or security-critical software, even frontier models generate highly inefficient, ugly code that works on the surface but fails under rigorous standards. Keeping the AI on a short leash allows developers to guide the model through these training gaps.
Which tools are best for implementing supervised AI generation?
Implementing a strict, human-in-the-loop workflow requires choosing platforms that offer granular control and transparent diff previews. Developers looking to integrate these practices can compare the top options in our guide to the best AI coding tools, which evaluates which platforms support robust manual permissions over fully autonomous execution. Slepak noted that his team successfully utilized their own custom review tools and a modified fork of an AI agent called Crush to maintain the necessary level of control over the past year.
Frequently asked questions
What is the Short Leash AI coding method?
It is a development methodology created by Greg Slepak of okTurtles that requires expert developers to remain actively in the loop, manually reviewing code diffs and approving every change made by an AI agent rather than letting it run autonomously.
Can beginners use the Short Leash AI method?
No. According to the creator, this method is specifically designed for professional, expert software developers whose skills already outclass frontier AI models in their area of expertise.
What is ‘YOLO’ mode in AI coding?
YOLO mode refers to setting an AI coding agent to run without human permissions, allowing it to automatically write and commit code without presenting diffs for human approval first.
To find platforms that let you maintain tight control over your codebase, check out our hands-on review of the best AI coding tools.
Best AI Coding Tools (2026): 7 Tested & Ranked →Source: Hacker News. Published July 3, 2026.
The AITW News Desk tracks model releases and AI product launches daily. Every story is fact-checked against its primary source before publishing and edited by Ali Zayed — and always links back to the original source.
AI Tools Worth is independent and unsponsored. Some linked guides contain affiliate links — they never change our verdicts.
Hands-on verdicts, real price changes and the launches that matter. No hype, no spam — unsubscribe anytime.