Crux: A Markdown-Native Multi-Agent Workflow Framework

Crux is a Markdown-native multi-agent workspace framework. Its core feature is defining agents, skills, and workflows using pure Markdown files without binding to a specific runtime environment. It aims to solve problems such as complex configuration, steep learning curves, and difficult version control in existing multi-agent frameworks. With the concept of 'documentation as configuration', it lowers the entry barrier and maintenance cost of agent systems.

With the development of AI agent technology, building multi-agent collaboration systems has become increasingly common, but existing frameworks have many configuration issues:

• Steep learning curve: Need to master specific configuration syntax and models
• Difficult version control: Structured configuration files are hard to diff and review
• Poor readability: Configuration is separated from documentation; understanding the system requires cross-file reference
• Vendor lock-in: Dependent on specific runtimes, leading to high migration costs

Crux takes 'Markdown as configuration' as its core philosophy. It writes all agent definitions, skill descriptions, and workflow orchestrations into .md files, which can be read without a special runtime. Its architecture uses a three-layer abstract model:

1. Agent Layer: Each agent is a Markdown file containing roles, capabilities, etc. (both documentation and configuration)
2. Skill Layer: Skills are Markdown files describing input/output, steps, etc., which can be reused
3. Workflow Layer: Orchestrates agent collaboration modes, supporting conditional branches, parallelism, etc. Declarative configuration brings advantages: human-readable, version-friendly, rich tool ecosystem, and progressive adoption.

Crux is suitable for various scenarios:

• Research team knowledge management: Use agents to define research directions, skills to define shared tools, and workflows to define research processes, allowing new members to quickly understand the working methods
• Customer service agent orchestration: Business personnel can directly edit Markdown to adjust the collaboration process of agents for intent recognition, knowledge retrieval, etc.
• Content creation pipeline: Build an automated workflow for topic selection, research, writing, editing, and publishing, reducing the cost of context switching.

Markdown Extension Strategy

Crux extends Markdown's expressive power through the following methods: YAML frontmatter metadata, convention-based tags, custom directives, link semantics

Execution Model

Supports static generation (compiled to JSON/Python), interpretive execution (directly reading Markdown), and hybrid mode

Comparison with Existing Frameworks

Feature	Crux	LangChain	AutoGen	CrewAI
Configuration Format	Markdown	Python Code	Python Code	Python/YAML
Runtime Dependency	None	Required	Required	Required
Readability	High	Medium	Medium	Medium
Version Control	Friendly	Average	Average	Average
Learning Curve	Low	High	High	Medium
Crux is more suitable for configuration-driven, documentation-first scenarios, while other frameworks are better in programming flexibility.

Crux is not a one-size-fits-all solution and has the following limitations:

• Ceiling of expressive power: Complex logic may be clumsy to express in Markdown
• Performance overhead: Parsing Markdown is slower than structured formats
• Tool maturity: Toolchain and debugging support are not as good as mature frameworks
• Community size: Fewer resources such as examples and plugins Selection should be based on scenarios; complex custom logic scenarios may be more suitable for traditional frameworks.

Future Directions

• AI-native documentation: Blurring the line between documentation and configuration, with natural language descriptions driving the system
• Collaborative development: Non-technical roles can participate in configuration, promoting cross-functional collaboration
• Explainable AI: Markdown definitions help understand AI decisions

Selection Recommendations

If you value configuration readability and version control, Crux is worth referencing; if you need complex custom logic and deep integration, traditional frameworks are more appropriate. Technical selection needs to balance constraints and understand differences in design philosophies.