# AI-Driven Development Harness

> End-to-End Software Delivery Automation with 7 Specialized Agents — Spec-to-Verified Code in 4 Automated Stages

**Tags:** Claude Code, Multi-Agent, TDD, Jira Integration, TypeScript

## Situation

A development team needed to ensure AI-generated code consistently met specifications across a multi-repository codebase. Manual code review was catching spec deviations too late, and the feedback loop between ticket creation and verified implementation was slow and error-prone.

## Task

Design an end-to-end development automation system that decomposes user stories into implementable subtasks, implements code using test-driven development, reviews it from both product and technical perspectives, and verifies the complete implementation matches the original user story.

## Actions

### 1. Stage 1: Story Decomposition

Automated story analysis with triage agent evaluating relevance across all repositories.

- Sub-agents investigate each repository in parallel — analyzing file structure, existing patterns, type definitions
- Impact analysis agent identifies affected systems and presents implementation options
- Engineer approves approach before subtask breakdown begins
- Generates structured subtask descriptions using separate front-end and back-end templates

### 2. Stage 2: TDD Implementation

Per-subtask implementation with strict RED → GREEN → REFACTOR quality gates.

- Test case design agent derives test cases organized by perspective (API, mapping, interactions, errors)
- Implementation agent writes production code within strict scope constraints
- Refactoring agent improves code quality while maintaining GREEN state
- Automated scope verification via diff analysis prevents unintended changes

### 3. Stage 3: Dual-Track Review

Parallel review by two specialized agents covering spec compliance and code quality.

- Spec review agent: specification consistency, interface requirements, prohibited-change violations
- Tech review agent: security, coding standards, performance, existing code impact
- Results consolidated into unified report with severity levels (CRITICAL / WARNING / INFO)

### 4. Stage 4: Story-Level Verification

Cross-functional verification after all subtasks complete.

- Story verification agent checks user story achievement, inter-subtask consistency, front-end/back-end alignment
- Full test suite execution with implementation summary generation
- Engineer performs manual testing before final approval

## Key Design Decisions

- Human-in-the-loop at every critical gate — AI does not proceed without engineer approval
- File-based communication between agents — enabling auditability and reproducibility
- Project management tool is read-only for AI — all write operations performed by engineer

## Results

- Code output consistently meets specifications through the RED → GREEN → REFACTOR cycle
- Spec deviations caught during implementation rather than during manual review
- Dual-track review covers both "does it match the spec?" and "is the code good?" simultaneously
- Complete harness delivered from requirements gathering to production deployment

---

Next case study: [Autonomous Build Pipeline](https://kcode.dev/en/portfolio/autonomous-pipeline.md)

Part of the [K Code portfolio](https://kcode.dev/en/portfolio.md). Web version: https://kcode.dev/en/portfolio/dev-harness · Contact: kenishida@kcode.dev