Deliberate Agentic Development: A Structured AI-Assisted Development Workflow

This article introduces the Deliberate Agentic Development open-source project, a structured workflow framework for AI-assisted programming that emphasizes human-AI collaboration, process standardization, and agent management. It aims to address issues such as inconsistent quality and insufficient context understanding caused by the stateless nature of current AI programming tools, helping teams efficiently complete software development from planning to delivery.

Large language models excel in code generation, but their direct integration into development workflows yields suboptimal results. Developers face issues such as inconsistent quality of AI-generated code, repeated modifications due to lack of context understanding, difficulty tracking change history involving AI, and unclear boundaries in human-AI collaboration. The root cause is that most AI programming tools are "stateless"—they do not understand project backgrounds, maintain conversation history, or follow team norms. Each interaction is isolated, making it hard to form a coherent collaborative process.

The project's core philosophy is that AI-assisted development should be a designed, traceable, and auditable collaborative process, reflected in three aspects:

1. Structured Workflow: Divide into clear phases, each with specific inputs, outputs, and acceptance criteria, where AI agents play specific roles;
2. Human Oversight Mechanism: Key decisions require human confirmation; AI suggestions are marked as "drafts" and proceed to the next phase only after manual review;
3. Seamless Agent Management: Unified lifecycle management (creation, configuration, monitoring, decommissioning), with agents archived after task completion.

The project defines a complete development process consisting of six phases:

• Requirements Analysis: AI assists in extracting requirements, identifying risks, and suggesting solutions, outputs a structured requirements specification, and humans confirm the accuracy of requirements;
• Architecture Design: AI suggests component decomposition, technology selection, and dependency analysis, outputs architecture decision records and C4 model diagrams, and ensures compliance with team architecture principles;
• Task Planning: AI splits tasks, analyzes dependencies, and estimates workload; humans determine priorities and deadlines;
• Coding Implementation: AI generates initial code, test cases, and synchronizes documents, follows coding standards, and adopts an "agile" (small steps) approach;
• Code Review: AI performs static analysis, security scanning, and performance hints; reports serve as references for human review;
• Delivery & Deployment: AI generates release notes, prepares rollback plans, configures monitoring, and archives documents.

Agent Roles:

• Architect Agent: Focuses on high-level design, provides solution comparisons, and does not generate code;
• Developer Agent: Executes coding tasks, generates test cases, and follows specifications;
• Reviewer Agent: Performs quality checks and categorizes issue priorities;
• Documenter Agent: Synchronizes document updates and supports multiple formats.

Toolchain Integration:

• Version Control: Standardizes Git commit messages and provides branch strategy hints;
• Project Management: Synchronizes tasks with Jira/Linear and updates work hours;
• CI/CD: Analyzes build failures and tracks test coverage;
• Communication Tools: Generates Slack/Feishu message summaries and supports asynchronous reviews.

Teams can adopt the framework through the following path:

• Pilot Phase (1-2 weeks): Select small functional modules, go through all six phases to familiarize with the workflow and tools, and collect team feedback;
• Expansion Phase (1-2 months): Apply to regular development tasks, maintain high human review participation, and adjust agent roles and checklists;
• Maturity Phase (3-6 months): Become the team's standard practice, optimize AI prompts, and establish team-specific best practice documents.

The limitations of the current version include:

• Context Window Limitation: Global understanding of large codebases is constrained by the LLM's context length;
• Domain Knowledge Dependency: Highly specialized fields (e.g., core financial systems) require more manual validation of AI suggestions;
• Creative Tasks: In breakthrough architecture design, AI plays more of a documentation and organization role rather than a source of creativity.

Deliberate Agentic Development provides a pragmatic framework that does not pursue full automation but instead establishes a sustainable human-AI collaboration model. By clarifying workflows, human oversight points, and agent role design, it helps teams improve efficiency while maintaining quality and process control. Its core insight: The greatest value of AI programming tools lies in establishing traceable, improvable, and collaborative engineering processes.