de-agentic-workflow: An AI-Assisted Workflow Framework for Data Engineering Teams

This article introduces the open-source project de-agentic-workflow, a configuration-first AI-assisted workflow framework for data engineering teams. Through standardized configurations, layered agent orchestration, and human supervision mechanisms, it addresses consistency (e.g., prompt style, security rules) and safety issues when teams use AI tools at scale. It supports mainstream data stacks like Snowflake and Airflow, requires no additional infrastructure, and is easy to customize and version-control.

With the popularity of AI coding assistants like Claude Code and GitHub Copilot, data engineering teams face challenges: members of a 10-person team use AI tools inconsistently, including differences in prompt styles, security boundaries, and code quality assumptions. This "drift" leads to inconsistent code styles, audit gaps, knowledge silos, and impacts collaboration efficiency and code quality.

Configuration-First Design: The framework exists as configuration files (not an independent service), which AI tools read as context. Advantages include no additional infrastructure needed, version-controllable (stored in Git), easy to customize, and rapid iteration. Agent Orchestration System: The root orchestrator identifies intent and routes requests; 13 professional agents (e.g., data modeling, ETL pipelines) have clear responsibilities; policy inheritance (security rules, Git workflows, architectural decisions) ensures consistency. Layered Approval Model: Read operations (queries, analysis) are executed autonomously; write operations (data modification, code deployment) require human approval, balancing efficiency and security.

External Integrations: Connects to Jira (task management), Notion (knowledge base), Snowflake (data operations), Azure DevOps (CI/CD) via the MCP protocol; Git integration supports GitHub/Azure DevOps. Workflow Standardization: Branch strategies (main branch protection, naming conventions), PR review processes (number of reviewers, checklists), incident response (grading, escalation paths), migration processes (templates, rollback strategies).

Application Value:

• Team: Maintains consistency in AI assistance and avoids chaos;
• AI developers: Verifies the feasibility of the configuration-first architecture;
• Managers: Balances efficiency and security through layered approval. Limitations: Primarily oriented towards the Claude Code ecosystem; adaptation to other AI tools requires adjustments; the configuration-first design may be insufficient for complex orchestration scenarios.

Future Directions: Support more AI tools (e.g., GitHub Copilot Chat), visual configuration editor, agent performance optimization, cross-team configuration sharing. Summary: de-agentic-workflow provides a practical AI-assisted framework for data engineering teams. Through configuration-first, agent orchestration, and human supervision, it addresses consistency and safety issues when using AI at scale, making it worth referencing.