Agentic Workflow Kit: A Declarative AI-Driven Development Workflow Framework

Original Author/Maintainer: aryeko Source Platform: GitHub Original Link: https://github.com/aryeko/agentic-workflow-kit Publish/Update Time: 2026-06-10T23:45:40Z

agentic-workflow-kit is a tracker-driven, specification-first delivery framework that supports Claude Code and Codex plugins, enabling automated PR/merge strategy management through declarative configurations. It addresses the limited participation of AI coding assistants in complex development processes (requirements analysis, task decomposition, code review, etc.), transforming AI from a passive responder into an active participant in project execution.

With the popularity of AI coding assistants like Claude Code and GitHub Copilot, developers have grown accustomed to real-time code suggestions in IDEs, but AI participation remains limited in complex software development processes (requirements analysis, task decomposition, code review, test validation). Most existing AI tools use conversational interactions, which are efficient for simple tasks but struggle with complex projects involving multi-step collaboration, state tracking, and quality assurance. Developers need a systematic solution to deeply integrate AI capabilities into software delivery workflows. agentic-workflow-kit was designed for this purpose, proposing a "tracker-driven, specification-first" development paradigm to transform AI from a passive responder into an active participant.

Spec Layer

Developers define project goals, acceptance criteria, and constraints in a declarative way (structured format or natural language), following the "convention over configuration" principle. This provides clear evaluation criteria for subsequent execution, allowing AI to perform retrospective self-checks.

Tracker Layer

The system's state management center records project status, tasks, and blocking issues. It is AI-native (understandable and operable by AI) and uses an event sourcing pattern:

• Auditability: Complete history for easy retrospective review
• Fault tolerance: Recovery from any historical state
• Collaboration: Multi-AI agents collaborate based on a unified state

Execution Layer

Converts specifications into code changes, including:

1. Claude Code/Codex plugins: Translate tracker tasks into code edits
2. Autonomous orchestrator (optional): Automatically schedules tasks, handles dependencies, and coordinates multi-AI agents

Declarative PR/Merge Strategy

Code review rules are declared (e.g., via YAML configuration). When AI submits a PR, it automatically checks if the strategy is satisfied; if not, it informs about missing conditions or triggers automatic fixes.

Task Decomposition & Dependency Management

1. Intelligent decomposition: AI analyzes specifications to identify parallel/serial task units
2. Dependency graph construction: Automatically identifies dependencies and builds a DAG
3. Critical path optimization: Prioritizes scheduling of critical path tasks to shorten delivery time

Multi-agent Collaboration Mechanism

• Role division: Architect, developer, test engineer, etc.
• Message bus: Structured message communication to avoid context confusion
• Conflict resolution: Provides arbitration mechanisms for resource competition

Claude Code Deep Integration

Officially supported by Anthropic with seamless integration:

• Context awareness: Claude reads tracker status to understand progress
• Action feedback: Claude edits are automatically synced to the tracker
• Intelligent suggestions: Hints at potential risks and improvements based on specifications

GitHub/Codex Workflow

Supports GitHub Copilot plugins:

• Automatic PR generation: Generates complete PRs based on task descriptions
• Review assistance: Pre-checks common issues during PR phase
• CI/CD integration: Links with GitHub Actions for full automation

Typical Application Scenarios

1. Rapid prototyping: AI participates throughout from requirements to prototype
2. Legacy code refactoring: AI analyzes and formulates plans for phased execution
3. Document synchronization: Automatic document updates when code changes
4. Open-source project maintenance: Automatically handles Issue classification, PR initial screening, etc.

State Machine-Driven Task Lifecycle

Each task's state transitions: To-do → In progress → Under review → Testing → Completed, triggered by events to ensure predictable processes.

Incremental Specification Updates

Supports incremental specification updates. AI identifies the scope of change impact and automatically adjusts task priorities and dependencies.

Pluggable Execution Backend

The execution layer uses a plugin architecture. In addition to Claude and Codex, it can integrate other AI models or custom tools in the future.

agentic-workflow-kit represents a paradigm shift in software development: from "human-led, AI-assisted" to "specification-driven, AI-executed, human-supervised". It does not replace developers but frees humans from tedious processes to focus on defining problems, formulating strategies, and accepting outcomes. As AI capabilities improve, this workflow will become a standard practice for complex software development. For organizations looking to enhance team efficiency and reduce communication costs, it provides a ready-to-use starting point.