Skip to main content

[DRAFT] ADR-009: Rationale for the Proposed Agentic Workflow

Context​

The integration of AI agents as full-fledged squad members requires a workflow that balances speed with extreme safety. We currently propose the 3-level rule hierarchy and branch-based collaboration.

Proposed Decision​

The proposed workflow is built on three pillars:

  1. Pillar 1 (Atomic Tasks): Every complex change is broken down into small, verifiable chunks in task.md. This allows for fine-grained human review.
  2. Pillar 2 (Branch Isolation): Agents never touch main. This ensures that even if an agent makes a mistake, it is isolated in a feature branch.
  3. Pillar 3 (Pre-approved Intent): The implementation_plan.md acts as a contract. The agent does not write code until the human squad agrees on the strategy.

Scope of Implementation Plan Requirement​

The implementation_plan.md requirement (Pillar 3) applies to structural changes only to balance safety with development velocity.

REQUIRED for:​

  • New ADRs or modifications to existing ADRs
  • Changes to .agent/ structure (rules, skills, workflows)
  • New sections in Docusaurus sidebar configuration
  • Breaking changes to content structure or navigation
  • Database schema changes or API modifications

NOT REQUIRED for:​

  • Content updates within existing pages
  • Typo or grammar corrections
  • Translation additions
  • Asset updates (images, diagrams)
  • Metadata-only changes (e.g., updating last_modified)

Rationale​

This scope limitation prevents bureaucratic overhead for routine maintenance while maintaining rigorous oversight for architectural decisions. The definition of "structural change" aligns with changes that could:

  • Break existing functionality or navigation
  • Affect multiple stakeholders or systems
  • Require coordination across team members
  • Have compliance or security implications

Expected Consequences​

  • Safety: High barrier against accidental repository corruption.
  • Clarity: Total traceability of "why" a change was made and "who" approved it.
  • Friction: Slower iteration cycles compared to direct editing, which is an acceptable trade-off for WEP's "Quality First" methodology.