Write "executable instruction manuals" for Agents using SOP thinking. AGL enables domain experts to reliably "steer" general-purpose Agents using natural language, injecting expert preferences/biases and proprietary knowledge into task execution workflows.
Large language models are already "knowledgeable and intelligent," but their default output tends to be average and generic to accommodate everyone. Yet great Agents in enterprise scenarios need "bias and taste." A truly deployable Agent must deeply involve domain experts, embedding domain habits, implicit rules, and quality thresholds into the system. Meanwhile, stable execution of real-world tasks is still limited by: inconsistent expression styles, missing critical information, and insufficient awareness of model/tool boundaries.
To address this, we propose AGL (Agent Guidance Language): AGL uses natural language expression, with its overall framework drawing from the management concept of SOP (Standard Operating Procedures), which business experts are already familiar with. At the same time, AGL is an agentic-friendly specification that considers Agent adaptability as much as possible, greatly improving the stability of Agent task execution. Task descriptions written following the AGL standard are called: "AGL Manuals".
This repository provides:
- ✨ AGL Specification (draft): Minimum Viable Product (MVP) specification
- 🧩 Templates: Ready-to-use manual templates (with examples)
- 📚 Documentation: Design philosophy, glossary, style guide
- 🧪 Validation Suggestions: How to self-check manual quality and stability
This project is initiated by BISHENG. Community contributions to refine the standard are welcome.
- Copy the template:
/templates/agl-manual-template.md - Fill in your task, completing the three main sections: "Problem Overview, Required Resources, Step-by-Step Instructions."
- Place your manual in
/examples/and submit a PR.
/docs/AGL-Spec-v0.1.md— Specification/templates/agl-manual-template.md— Official template/examples/— Example manualsCONTRIBUTING.md— How to contribute (including AGL Proposal process — AGLP)LICENSE— MIT
MIT License