Agentic Development Workflow

A reusable, human-governed workflow for planning and executing analytics, data science, and applied AI projects with coding agents.

The methodology separates project definition from implementation through a structured handoff. PLANNING.md turns an initial idea into a bounded project brief; EXECUTION.md converts that approved brief into an operating plan with task-specific workflows, role separation, reproducibility requirements, and quality gates.

PLANNING.md  ->  PROJECT_BRIEF.md  ->  EXECUTION.md  ->  Project artifacts
 define it        lock the scope        build it          verify it

Design highlights

• Structured project intake: ten lock conditions prevent execution from beginning with an ambiguous goal, undefined inputs, or a vague definition of done.
• Proportional process: Minimal, Standard, and Research modes scale the documentation and controls to the size of the project.
• Task-specific execution: five workflows cover standard development, evaluation-driven research, debugging, quick changes, and planner-first work.
• Human decision ownership: scope, architecture, deployment, interpretation, and release decisions remain explicit human approval points.
• Independent review: judging roles can run in fresh contexts with limited inputs, reducing the risk of agents validating their own assumptions.
• Research traceability: experiment configurations, dataset and prompt versions, metrics, failure modes, and next actions are recorded as part of the workflow.

How it works

1. Define the project

Start a planning conversation with PLANNING.md. The protocol guides the discussion through the problem, users, inputs, outputs, non-goals, technical constraints, evaluation strategy, reproducibility, security, and definition of done.

The planning phase does not close until its lock conditions are satisfied and the user approves the resulting brief.

2. Create the handoff contract

The planning session produces PROJECT_BRIEF.md in a fixed schema. This file becomes the source of truth for the build and records:

• The project goal and intended users
• Explicit non-goals and technical boundaries
• Inputs, outputs, and deployment expectations
• Evaluation and reproducibility requirements
• Security constraints and testable completion criteria

If implementation reveals that the brief is incomplete or incorrect, the workflow requires a documented, human-approved amendment instead of silent scope drift.

3. Execute and verify

Provide the approved brief and EXECUTION.md to a coding agent. The operating standard selects the appropriate project mode and workflow, then defines how the work should be planned, implemented, tested, reviewed, documented, and approved.

The execution standard includes:

• Standard, evaluation-driven, debugging, quick-task, and planner-first workflows
• Planner, implementer, tester, reviewer, security reviewer, debugger, experiment analyst, and documentation roles
• Templates for specifications, plans, task tracking, reviews, and experiment logs
• Quality gates covering scope, correctness, testing, security, reproducibility, and documentation

Repository contents

File	Purpose
README.md	Explains the methodology, design principles, and recommended workflow.
PLANNING.md	Conducts the intake conversation and produces an approved project brief.
EXECUTION.md	Defines execution workflows, agent roles, project templates, and quality gates.

Using the workflow

1. Begin a planning session with PLANNING.md and describe the proposed project.
2. Review and approve the generated PROJECT_BRIEF.md.
3. Place the approved brief in the target project repository.
4. Start the build with EXECUTION.md and the approved brief.
5. Retain human approval at the planning, architecture, review, interpretation, and release gates.

PLANNING.md and EXECUTION.md are reusable methodology files. The generated brief and project-specific artifacts belong in the project being built.

Scope

This is a vendor-neutral Markdown methodology rather than an installable framework. Its controls are designed for human-directed analytics, data science, and applied AI projects and can be used with any coding agent capable of following file-based instructions.

The workflow has been developed through use on individual prototypes, evaluation studies, and academic projects. It defines practical operating controls but does not claim automated enforcement or team-scale validation.