Applying AI-DLC to Team Projects: The ARMIQ Case, Extended with Subagents and Custom Skills

AI-DLC is a development approach in which an AI Agent handles everything from requirements analysis to design, code generation, and review. ARMIQ applied the AI-DLC ToolSet simultaneously to two projects, AETL and ACMS, enabling parallel development. AETL was a Brownfield Node.js monolith (Express + React), while ACMS was a Greenfield TypeScript-based Lambda/Aurora PostgreSQL serverless architecture.

AETL comprised 8 Units, 12 US and 17 TT, and 16 test files, while ACMS set a target of 5 Units, 31 US, 55 APIs, and 95% coverage. AETL focused on improving the security and structure of an existing data pipeline, while ACMS involved newly designing a standards-based server that sends documents to AWS S3 for unified management.

The AI-DLC workflow follows an INCEPTION → CONSTRUCTION → OPERATIONS structure, but stages are always divided into items that always execute and items that execute conditionally. Workspace Detection, Requirements Analysis, Workflow Planning, Code Generation, and Build and Test are ALWAYS stages, Reverse Engineering runs only for Brownfield, and User Stories, Application Design, and Units Generation are attached depending on complexity. This Adaptive Workflow, which varies its path according to state, combined with Unit decomposition, allowed Brownfield and Greenfield to be handled the same way.

Orchestration was handled by the Main Agent, while analysis and generation were delegated to specialized Subagents. Planning is performed by the Main Agent and, once user approval is obtained, the Subagent implements only the approved Plan. The Reverse Engineering Subagent produces 8 artifacts ranging from business-overview.md to code-quality-assessment.md, and the Code Generation Subagent generates code per Unit; the code is reflected directly into the workspace root, while design outputs are stored under aidlc-docs/construction/{unit-name}/code/. Subagent responses were constrained to a fixed schema — STATUS, UNIT, FILES, TESTS — to reduce the Main Agent's context consumption.

• context7: Up-to-date library/framework documentation
• aws-knowledge-mcp-server: Official AWS reference
• tavily: Web research during the INCEPTION stage

All three servers were connected to the Main Agent, while only context7 and aws-knowledge-mcp-server were connected to the Subagents. Generated code was reviewed again with a dedicated per-Unit Code Review Skill, adding a quality guardrail.

• requirements-generator: Generates requirements/constraints documents from PDFs, markdown, and images
• git-merge: Classifies and handles conflicts in aidlc-state.md and audit.md as state-file conflicts versus code conflicts

The requirements-generator does not invent content that isn't in the documents, and leaves unverifiable items marked as lacking information. git-merge routes merge conflicts arising after parallel Unit work to either automatic merging or user confirmation, sustaining a flow like ACMS's, where a 4-person team split 5 Units into Waves for parallel work.