Agent Lab: A Markdown-Driven AI Agent OS for Professional Work

Agent Lab is a document-first reusable AI agent operating system designed specifically for professional work scenarios. It does not build web applications, store sensitive customer data, or bind to specific AI platforms. Its core goal is to make agent behaviors clear, portable, testable, secure, and easy to continuously improve. The project treats agents as reusable operational programs, solving practical problems such as agent selection, information requirements, integration adaptation, security checks, deliverable forms, evaluation, and version control, suitable for technical development and project delivery.

Agent Lab adopts a "document-first" design philosophy, with the core goal of making agent behaviors clear, portable, testable, secure, and easy to continuously improve. Its core idea is: professional AI work should treat agents as reusable operational programs rather than one-time prompts. This transformation helps solve key problems in practical work: Which agent is suitable for this task? What information does the agent need before starting work? Which integration adapters should be used? What security checks need to be applied? What form should the final deliverable take? How to evaluate and version-control agent behaviors?

Agent Lab uses a modular layered architecture, including the following core components:

• Agent Definition (agents/): Reusable role definition files, including mission, workflow, boundaries, guardrails, etc.
• Skill Library (skills/): Reusable capability modules, including purpose, input requirements, execution process, etc.
• Workflow Orchestration (workflows/): Repeatable execution processes covering scenarios such as delivery and analysis.
• Prompt Templates (prompts/): A collection of prompts to activate agents/skills/workflows.
• Output Templates (templates/): Structured output format definitions to ensure deliverable consistency.
• Integration Adapters (integrations/): Platform-agnostic connection methods for external tools.
• Security and Governance: Includes guardrails (security rules), governance (ownership/policies), threat-models (threat models), secrets-management (secret management).
• Observability and Evaluation: Includes operation logs, decision records, evaluation strategies, etc.

Agent Lab prioritizes security and implements multi-layer protection:

• Default Read-Only Principle: All agents are read-only by default; explicit approval is required for writing/deployment/release/sensitive data usage.
• Hierarchical Access Control: Five levels of permissions (0-1: read-only, 2: draft/write, 3: deploy/execute, 4: sensitive data access); higher levels require explicit approval.
• Data Protection Specifications: Prohibit storing secrets on GitHub; desensitize/anonymize sensitive data; manage secrets via environment variables; prohibit pasting secrets in prompts/examples; prohibit storing original customer files in the repository.

Agent Lab is suitable for various professional scenarios:

• Customer Website Delivery Process: Collect requirements → Create specifications → Generate build prompts → Build MVP → Review → Deploy → Test → Iterate.
• Technical Evaluation and Research: Supports library evaluation, AI-assisted MVP creation, technical document writing and self-review.
• Multi-Agent Collaboration: The orchestration directory provides routing, handover, and multi-agent workflow guidance, supporting complex task collaboration.

Standard Usage Flow:

1. Select an agent from agent-registry.md
2. Review the agent file (mission, input, boundaries, etc.)
3. Select the corresponding workflow (if the task is repeatable)
4. Activate the agent/workflow using prompts
5. Use templates for structured output
6. Check quality against examples
7. Apply guardrails/contracts/threat models before performing risky work
8. Record decisions and results

Extension Customization:

• Create a New Agent: Copy the template → Define mission/input/workflow, etc. → Add risks/guardrails/evaluation → Update the registry.
• Create a New Skill: Copy the template → Define purpose/process → Link to agents/workflows → Add prompts/examples.

Version Management: Uses semantic versioning: MAJOR (structural/breaking changes), MINOR (new components added), PATCH (optimizations/fixes). Agent files have independent versions, and changes are recorded in the changelog.

Project Value:

1. Standardize AI workflows to ensure consistency
2. Reduce security risks and protect sensitive data
3. Improve collaboration efficiency and support team collaboration
4. Ensure delivery quality through the evaluation system
5. Achieve knowledge precipitation and become reusable organizational assets

Agent Lab provides a framework for teams/individuals to systematically integrate AI into professional workflows.