This repository provides AGENTS.md, a single configuration file designed to transform AI coding assistants from inexperienced interns into senior engineers. It addresses common issues like sycophancy, unnecessary refactoring, and lack of code verification by synthesizing established principles for LLM coding. The primary benefit is a significant improvement in the reliability and quality of AI-generated code across various platforms.

How It Works

The core approach involves a drop-in AGENTS.md file that agents read to modify their behavior. It synthesizes Andrej Karpathy's principles on LLM coding failure modes and Boris Cherny's Claude Code workflow. Key behavioral changes include agents pushing back on incorrect user input, making minimal necessary changes, writing and running verification steps before completion, and surfacing ambiguities rather than guessing. The file contains a fixed behavioral scaffold (sections 0-9) and two editable sections: "Project context" (section 10) for technical details, and "Project Learnings" (section 11) where the agent logs corrections, creating a compounding memory.

Quick Start & Requirements

The easiest installation involves fetching AGENTS.md from the repository and saving it as ./AGENTS.md in your project root. For specific agents like Claude Code and Gemini CLI, symbolic links (CLAUDE.md, GEMINI.md) should be created pointing to AGENTS.md (or the file copied if symlinks are unavailable). Users must then fill in Section 10 ("Project context") with verified details like stack, build commands, and directory layout, estimated to take about five minutes. No other non-default prerequisites are mentioned.

Highlighted Details

• Cross-Agent Standard: AGENTS.md is an open standard supported natively by agents like Codex, Cursor, Copilot, Devin, Aider, and others. Claude Code and Gemini CLI are supported via specific filenames (CLAUDE.md, GEMINI.md).
• Behavioral Enforcement: Agents are prompted to verify code before committing, push back against user errors, and limit changes strictly to the requested scope.
• Compounding Memory: Section 11 automatically logs agent mistakes and corrections, allowing the agent to learn and improve over time, acting as a "trained reflex."
• Minimalist Design: The file is intentionally kept concise (~200 lines) to ensure rules remain loaded within agent context windows.

Maintenance & Community

The project is stewarded by the Linux Foundation's Agentic AI Foundation, indicating a commitment to its development as an open standard. Specific community links (Discord/Slack) or details on other contributors are not provided in the README.

Licensing & Compatibility

The project is licensed under the MIT license, permitting free use, modification, and distribution, including for commercial purposes.

Limitations & Caveats

For very large codebases, AGENTS.md might exceed typical context limits, requiring advanced sharding techniques using file imports or scoped rule directories, though the README suggests a single file remains optimal for most users. The effectiveness and specific implementation details may vary slightly between supported AI agents. Initial setup requires manual configuration of the AGENTS.md file and potential symlinks.