Doctrine: Python-style DSL and Compiler Architecture for AI Agent Instructions and Workflows

This article provides an in-depth analysis of the Doctrine project, whose core is to design a Python-style domain-specific language (DSL) and compiler architecture for AI Agent development. It aims to address engineering challenges in Agent development (such as prompt engineering dilemmas and workflow definition pain points) by offering three core abstractions: instructions, workflows, and contracts, and compiling to generate standardized AGENTS.md documents, thereby enhancing the reusability, maintainability, and testability of Agent development.

With the improvement of LLM capabilities, AI Agents are moving towards production deployment, but they face two core challenges:

1. Prompt Engineering Dilemma: Plain text prompts are difficult to version control, lack structure, are hard to maintain, and pose great testing challenges;
2. Workflow Definition Pain Points: Hardcoding lacks flexibility, configuration files have limited expressiveness, and visual editors are not suitable for complex logic—balancing maintainability and expressiveness among these three is difficult.

Doctrine proposes an innovative approach: designing a Python-style DSL and compiler to generate AGENTS.md.

• Reasons for Choosing Python-like DSL: Wide developer base, strong readability, balanced expressiveness and conciseness, and ability to leverage existing toolchains;
• Three Core Abstractions:

1. Instructions: Structured definition of Agent behavior rules and capabilities, supporting modularity and type safety;
2. Workflows: Declarative definition of multi-step task flows, improving readability and modifiability;
3. Contracts: Formalized Agent interaction protocols, drawing on contract programming to ensure interaction consistency.

Doctrine compiles the DSL into AGENTS.md (Markdown format):

• Advantages of AGENTS.md: LLM-friendly, human-readable, rich tool ecosystem, version control-friendly;
• Compiler Process:

1. Lexical Analysis: Convert character streams into Token sequences;
2. Syntax Analysis: Build an Abstract Syntax Tree (AST);
3. Semantic Analysis: Check the correctness of types, references, and constraints;
4. Code Generation: Render Markdown templates to generate AGENTS.md.

Doctrine brings significant engineering value to Agent development:

• Reusability: Supports import inheritance, library ecosystem building, and semantic version management;
• Testability: Static analysis, simulation execution, regression testing;
• Maintainability: Single source of truth, documentation as code, IDE support;
• Collaboration-Friendly: Code review, division of labor and collaboration, knowledge precipitation.

Applicable Scenarios: Enterprise-level Agent platforms, Agent markets/stores, multi-Agent systems, compliance-sensitive scenarios; Limitations: Developers need to learn a new DSL, the toolchain ecosystem is yet to mature, expressiveness needs to be balanced, and it depends on a specific runtime.

Future Directions: Promote AGENTS.md standardization, deep IDE integration, visual editors, multi-runtime support, AI-assisted development; Conclusion: Doctrine introduces best practices in software engineering (DSL, compilation principles, etc.), providing an engineering solution for Agents from prototype to production, and is an important direction for AI Agent infrastructure innovation.