Plan-Based Agentic Workflow: A Plan-First Agent Development Workflow

This article introduces the Plan-Based Agentic Workflow—a collection of agent skills supporting a plan-first development process, with the core principle of conducting requirement analysis and task planning before coding. It includes three key skills: feed-pm (requirement analysis and task drafting), implement-pm (task implementation), and review-pr (PR review), which form a complete closed loop from requirements to code, balancing AI automation and human control. Original author: mvagnon; Project source: GitHub (link: https://github.com/mvagnon/plan-based-agentic-workflow); Update time: 2026-05-30T03:45:39Z.

Many current AI-assisted development tools overfocus on code generation, which easily leads to code deviating from real requirements or ignoring edge cases. The plan-first model draws on traditional software engineering best practices: requirement clarification (ensuring shared understanding), task decomposition (splitting into manageable units), reviewability (clear outputs at each stage), and traceability (complete link from requirements to code). For agents, it also reduces context window pressure, as each stage only focuses on current information.

feed-pm is the starting point of the workflow. After receiving user requirements, it performs: 1. Repository analysis (understanding codebase structure, tech stack, and existing patterns); 2. Requirement decomposition (breaking down high-level requirements into specific technical tasks, considering dependencies, priorities, and risks); 3. PM task drafting (generating detailed documents including task description, acceptance criteria, workload estimation, relevant file locations, potential risks, and mitigation measures). The task document can be manually reviewed and adjusted to ensure the correct direction.

implement-pm: Executed after task approval: creates an independent branch (to isolate work), generates code according to specifications (aligning with project style), automatically drafts a Draft PR (including implementation summary and test instructions). It only processes approved tasks to ensure controllability. review-pr: Automated PR review: checks specification compliance (whether it deviates from the task), code quality (style, bugs, security, performance), and generates a structured review report (compliant parts, parts needing modification, suggestions, merge evaluation). Qualified PRs can be marked as Ready for Review to enter the manual process.

The three skills form a complete loop: 1. User submits requirements → feed-pm generates tasks; 2. Manual review and approval of tasks; 3. implement-pm implements and creates PR; 4. review-pr automatically reviews PR; 5. Manual final review and merge; 6. New requirements enter the loop. This design retains human control at key decision points, while agents take on tedious analysis and implementation work.

The project aims for cross-platform compatibility. Skill definitions use a universal format and support multiple agent runners: Codex (OpenAI), Claude Code (Anthropic), Antigravity CLI, etc. Runner-specific metadata is optional; unsupported runners can ignore it. Core workflow instructions are universal, ensuring skills can be reused across different environments without rewriting.

Recommendations for team adoption: 1. Start small (try with small features or bug fixes); 2. Establish a review culture (manual review of tasks generated by feed-pm is key to quality); 3. Iteratively optimize prompts (adjust skill system prompts based on the project); 4. Integrate with existing tools (combine with CI/CD and project management tools without replacing existing processes).

The Plan-Based Agentic Workflow is a pragmatic path for AI-assisted development. It does not pursue full automation but leverages human-agent collaboration to发挥各自 strengths. Compared to direct AI code generation: it adds planning and review stages, which may seem slower but reduces rework, improves maintainability, and enhances controllability. Compared to manual processes: automation assistance allows developers to focus on high-value work. It is suitable for teams exploring AI-assisted development, providing a structured and flexible starting point.